The Cryptonomist

Every week, David Pierce distills the internet’s best tech, entertainment, and app discoveries into a single, opinionated guide — and Installer No. 136 from The Verge might be one of the more eclectic editions yet. From an unexpectedly great film to fresh app updates and a gadget Pierce admits he loves without knowing why, this week’s edition covers a lot of ground in the most personal way possible.
Key takeaways
Installer No. 136, published July 18, 2026, is David Pierce’s weekly technology and entertainment recommendations guide at The Verge.
Pierce calls The Odyssey the movie of the summer — surprising even himself after going in with low expectations.
The edition highlights a notable update to a popular note-taking app and a new app built for organizing photos.
Pierce is currently recording the next season of the Version History podcast, with the current season’s finale dropping on Sunday.
Readers can submit their own picks and recommendations directly to installer@theverge.com.
What Installer No. 136 is actually about
The Installer newsletter has built a quiet, loyal following at The Verge by doing something most tech publications don’t bother with: sounding like a real person. Pierce opens each edition with a loose personal update — what he’s reading, playing, watching, or fiddling with that week — before rolling into the actual recommendations. That mix of intimacy and curation is the whole point.
This week’s personal inventory is characteristically wide-ranging. Pierce has been deep in production on the next season of Version History, The Verge’s podcast — with the current season’s finale landing on Sunday. He’s also been exploring new Knockout Tour routes in Mario Kart World, setting up a Flipper Busy Bar (which he describes as something he loves but hasn’t found a use for yet), and going down rabbit holes on data center heists, the origin story of Calvin and Hobbes, and an episode of Revisionist History that apparently taught him more about Staten Island than he ever intended to know.
There’s also a detour through the history of the very first chatbot and some reading on Backyard Baseball. It’s the kind of week that would feel overwhelming in a meeting but reads effortlessly in a newsletter.
The Odyssey: A standout recommendation
The headline recommendation this week is The Odyssey, and Pierce doesn’t bury the lead. “I’ll be honest: I expected this movie to not be great,” he writes. The reasoning was sound enough — an ambitious story and the honest acknowledgment that not every ambitious film connects. But his verdict landed hard: wrong.
Calling it the movie of the summer is a significant stamp from someone who covers entertainment professionally. Pierce came in skeptical and left converted, which is arguably a stronger endorsement than enthusiasm from someone who was already sold on the premise. For anyone weighing whether to see it, that kind of reluctant praise tends to travel further than a glowing review from a fan.
App Updates and New Tools Worth Knowing
A note-taking app gets better
Pierce flags a meaningful update to a popular note-taking app this week — enough to earn a dedicated mention in the Drop section. Note-taking apps sit in a competitive, often overhyped corner of the productivity space, so a genuine improvement worth calling out is more useful signal than most app-store release notes tend to provide.
A new way to organize your photos
Alongside the notes update, Pierce spotlights a new photo organizing app. Photo management is one of those perpetual problems that every platform promises to solve and few actually do — so a fresh option aimed specifically at organization rather than editing or sharing is worth tracking, especially if it earns a spot in an edition of Installer.
What makes these mentions meaningful in context is that Pierce isn’t running a traditional review column. He’s surfacing things he’s personally used or found interesting that week. That lowers the volume of what gets through and raises the implied reliability of what does.
Reader Engagement and the Community Behind the Newsletter
One consistent feature of Installer is that Pierce treats it as a conversation rather than a broadcast. Each edition closes with an open invitation: tell him what you’re reading, watching, playing, listening to, or soldering together. The contact address is installer@theverge.com, and he actively frames reader input as the best part of the whole thing.
That posture matters more than it might seem. In a media environment where most newsletters are one-directional by design, a columnist who genuinely routes his own discovery process through reader suggestions creates a feedback loop that makes the product more durable over time. The tips he receives presumably feed future editions, which makes the engagement ask feel like something more than a boilerplate CTA.
If there’s a broader point embedded in what Installer does well, it’s this: the newsletter succeeds not because it covers the most ground, but because it covers ground that someone actually covered first. Pierce’s credibility comes from the specificity of his taste, and that specificity only holds up as long as the recommendations stay genuinely personal. Edition 136 clears that bar without much difficulty — and for a weekly column now well past its 130th entry, that consistency is harder to maintain than it looks.
FAQ
What is Installer No. 136 about?
Installer No. 136 is a technology and entertainment guide by David Pierce at The Verge, featuring personal app picks, gadget discoveries, and entertainment recommendations for the week of July 18, 2026.
Which movie does David Pierce recommend in this edition?
David Pierce highlights The Odyssey as an unexpectedly excellent film — calling it the movie of the summer after initially going in with low expectations.
Are there any app updates mentioned?
Yes. The edition mentions a notable update to a popular note-taking app and a new app designed specifically for organizing photos.
How can readers share their ideas with David Pierce?
Readers can send their ideas, tips, and recommendations directly to installer@theverge.com.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Take-Two Interactive has locked in November 19, 2026 as the launch date for Grand Theft Auto VI, with CEO Strauss Zelnick framing the release as a defining moment for the company’s financial future — and the numbers backing that confidence are hard to argue with. In a letter to shareholders ahead of the September annual meeting, Zelnick described fiscal year 2027 as a potential “major inflection point,” with GTA VI and Take-Two’s existing catalog carrying the weight of that projection.
Key takeaways
Grand Theft Auto VI is confirmed to launch on November 19, 2026, with Take-Two describing the date as “planned” in its SEC filing.
Fiscal 2026 Net Bookings reached $6.72 billion, beating guidance by $750 million, with net revenue at $6.66 billion.
Recurrent consumer spending — post-launch purchases — hit $5.20 billion, representing 78% of total revenue.
Adjusted EBITDA of $1.4 billion blew past the $919.5 million target, triggering maximum executive bonuses.
Take-Two forecasts operating cash flow above $1 billion for fiscal 2027, tied heavily to the November launch.
GTA VI and Take-Two’s Most Important Launch in Years
The confirmation itself carries a small asterisk. Zelnick used the word “planned” in his shareholder letter — language typical of SEC filings, where executives avoid absolute commitments. As observers noted, that phrasing reflects legal caution rather than genuine scheduling doubt. There is currently no evidence of any change to the November target.
What the letter does make clear is how much is riding on the launch. Zelnick pointed to operating cash flow above $1 billion as the fiscal 2027 target, with GTA VI as the primary engine. That figure notably comes without specific booking targets attached, leaving analysts to weigh the ambition of the projection against the execution variables still in play — including consumer reception to the game’s pricing and disc-free format, which generated a mixed reaction when first announced.
GTA Online and GTA+ Keep the Engine Running
One of the less-discussed parts of Zelnick’s letter is how well Rockstar’s existing live-service ecosystem is performing. GTA Online continues to outpace expectations, with the December 2025 update A Safehouse in the Hills singled out as one of the most successful seasonal drops to date. The GTA+ subscription service has also grown its membership base, helped by perks that now include NBA 2K26 in its library.
That matters strategically. If GTA VI’s launch faces any friction — whether from pricing sensitivity or a broader market reaction — the existing GTA Online ecosystem provides a significant revenue floor. The live-service model has effectively turned a game released over a decade ago into a consistent cash generator, something few franchises in any entertainment category can claim.
Take-Two’s Fiscal 2026 Financial Highlights
The raw numbers from fiscal 2026 make a strong case that Take-Two arrived at this GTA VI launch window from a position of real financial health, not desperation.
Record Net Bookings and Revenue Breakdown
Take-Two closed fiscal 2026 with Net Bookings of $6.72 billion — a non-GAAP measure that tracks signed orders and contracts — landing $750 million above its own guidance. Total net revenue came in at $6.66 billion. The revenue split was remarkably even: console and PC sales contributed $3.32 billion, while mobile revenue came in just slightly higher at $3.33 billion.
The deeper story is in recurrent consumer spending. Post-launch purchases — subscriptions, virtual currency, season passes — reached $5.20 billion, accounting for 78% of total revenue. That figure illustrates how completely the industry has shifted toward ongoing monetization rather than one-time sales. For Take-Two, it also means the company has built structural revenue that doesn’t depend on any single title’s launch window.
Grand Theft Auto V and Red Dead Redemption 2 Still Selling
Grand Theft Auto V has now sold nearly 230 million units since its original release — a number that defies easy comparison in entertainment history. Red Dead Redemption 2 has crossed the 80 million units mark. Both titles continue to generate revenue more than a decade and six years after launch, respectively, which speaks to the durability of Rockstar’s IP and the effectiveness of Take-Two’s long-tail monetization strategy.
Adjusted EBITDA and Executive Bonuses
Fiscal 2026 adjusted EBITDA — excluding interest, taxes, depreciation, and amortization — hit $1.4 billion. That figure surpassed the internal target of $919.5 million by a wide margin and automatically triggered maximum executive bonuses under the company’s compensation structure. The scale of the beat suggests the fiscal year performed significantly ahead of internal modeling, not just public guidance.
Fiscal 2027 Outlook and Governance
Operating Cash Flow Forecast Above $1 Billion
Zelnick’s forecast of operating cash flow exceeding $1 billion in fiscal 2027 is the headline number for investors thinking beyond the November launch. The projection signals confidence, but the absence of specific booking targets for the year creates a wider range of outcomes than investors might prefer. The operating cash flow forecast hinges significantly on GTA VI’s commercial performance in its first few months on shelves.
Take-Two shares closed at $239.57 on July 16, 2026, up nearly 13% over the prior month. The stock still sits below its 52-week high of $265.94 — a peak that came before the GTA VI pricing debate and disc-free format announcement drew pushback from parts of the player base. The gap between that high and current levels reflects an investor community that is broadly optimistic but not yet fully convinced.
Shareholders to Vote on Governance and Executive Pay
Take-Two’s virtual annual meeting is set for September 17, 2026. Shareholders will vote on the election of 10 directors, a non-binding say-on-pay resolution, a proposal to amend the company charter to limit certain officer liability under Delaware law, and ratification of Ernst & Young as auditor. The timing places the governance vote roughly two months before the GTA VI launch — meaning any organizational instability from a contested vote would land at a particularly sensitive moment.
There’s also a broader industry context worth noting. The question of whether the $79.99 price point — and a digital-only launch format — will hold up under real consumer pressure remains unanswered. Sony faced a similar test with its own digital-only pivot and weathered a significant backlash. Whether nearly a decade of built-up demand for GTA VI is enough to absorb pricing friction is the single variable no financial model can fully account for. An August 7th earnings call may offer the first hard data point, with speculation that Take-Two could reveal pre-order figures or additional marketing at that event.
FAQ
When will Grand Theft Auto VI be released?
Grand Theft Auto VI is confirmed to launch on November 19, 2026, according to Take-Two’s shareholder letter filed with the SEC.
How did Take-Two perform financially in fiscal 2026?
Take-Two reported record fiscal 2026 Net Bookings of $6.72 billion, total net revenue of $6.66 billion, and adjusted EBITDA of $1.4 billion — all significantly above guidance and internal targets.
What is the outlook for Take-Two’s fiscal 2027?
Take-Two forecasts operating cash flow above $1 billion for fiscal 2027, with GTA VI identified as the primary revenue catalyst for the year.
What governance issues will shareholders vote on in 2026?
At a virtual annual meeting on September 17, 2026, shareholders will vote on director elections, an executive pay resolution, a charter amendment to limit officer liability under Delaware law, and ratification of Ernst & Young as the company’s auditor.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Apple is about to give its smallest tablet the most significant iPad mini upgrade in five years — and the centerpiece isn’t just a new chip. According to supply chain reports and Bloomberg’s Mark Gurman, the next iPad mini will ditch its LCD panel entirely and adopt OLED technology for the first time. That’s a meaningful shift for a device that has largely stayed in the shadow of its bigger siblings.
Key takeaways
The next iPad mini is expected to feature an OLED display for the first time, replacing the current Liquid Retina LCD panel.
Apple plans to upgrade the chip from the A17 Pro to the A19 Pro, a jump that should noticeably improve performance in demanding tasks.
The refresh rate is rumored to remain at 60Hz — no ProMotion, unlike the iPad Pro’s 120Hz display.
The new iPad mini is expected to launch as early as October 2026 with a starting price of $699, up $100 from the current model.
A refreshed base iPad with an updated processor is also in the pipeline, though not until early 2027.
Apple’s Major iPad Mini Upgrade for 2026
The last time the iPad mini saw a redesign this substantial was 2021. What’s coming this fall, according to multiple reports, is a device that would look similar on the outside but perform and display content in a fundamentally different way.
OLED Display: What It Actually Changes
An OLED panel delivers deeper blacks, higher contrast ratios, richer HDR rendering, and potentially better battery efficiency compared to LCD. For a tablet used heavily for streaming, gaming, and reading, those aren’t marginal improvements — they’re the kind of changes you notice in the first five minutes of use.
Processor Upgrade from A17 Pro to A19 Pro
The chip story is just as compelling. The current iPad mini runs on the A17 Pro, which already supports Apple Intelligence. The incoming model is expected to carry the A19 Pro chip — two generations ahead — bringing stronger processing power that should make a real difference in tasks like video editing, photo processing, and AI-driven features. Gaming performance should benefit too.
This is where the 2026 iPad mini starts to feel like a different kind of product. More on that in a moment.
Performance and User Experience Enhancements
The combination of OLED visuals and a significantly faster chip repositions what the iPad mini is actually for. Until now, it has been primarily an entertainment tablet — easy to hold, light, good for consuming content. The A19 Pro changes that equation.
Broader Use Cases Beyond Entertainment
With the processing power of the A19 Pro, the iPad mini becomes a more credible tool for demanding workflows. Video editors who want a portable station, developers testing apps, or professionals who need Apple Intelligence features on the go will find a device that can keep up. The OLED screen makes that work more visually accurate too.
It’s a notable strategic shift for Apple. The iPad mini had carved out a niche as the “fun” tablet — easy to toss in a bag, great for a flight. Adding serious processing muscle without increasing the form factor means Apple is betting there’s an audience for a compact professional-grade tablet, not just a compact entertainment one.
Retained 60Hz Refresh Rate Despite OLED
Here’s the catch that will frustrate some buyers. Despite the OLED upgrade, the refresh rate is reportedly staying at 60Hz. This is a fixed 60Hz rate, not the adaptive ProMotion technology that gives the iPad Pro refresh rates up to 120Hz.
For most users, 60Hz on an OLED panel will still look noticeably better than 60Hz on an LCD — the pixel response and contrast make a visual difference even at the same frame rate. But users hoping for the buttery-smooth scrolling that ProMotion delivers are going to be disappointed. This is clearly where Apple draws the product line between the mini and the Pro.
Launch Timing and Pricing Details
Expected October 2026 Release and $699 Starting Price
According to Gurman’s Bloomberg reporting, the new iPad mini is expected to arrive as early as October 2026. The timing fits a pattern Apple has used before — fall hardware announcements ahead of the holiday shopping season.
The price, however, is climbing. The new model is expected to start at $699, which is $100 more than the current iPad mini’s price. As The Verge noted, Apple already raised prices across its product lineup earlier this year — including the current iPad mini — with Tim Cook attributing the increases to an ongoing RAM shortage. The OLED display and A19 Pro chip give Apple clear justification for the higher price point, but the $699 entry barrier is a meaningful step up from where the mini has traditionally sat in Apple’s lineup.
That price increase is worth watching closely. The iPad mini has always competed partly on value — a premium tablet experience at a fraction of the iPad Pro’s cost. At $699, the gap narrows, and buyers will need to weigh whether the OLED and chip upgrade justify the premium over the older model, which will likely see a price cut.
Upcoming Base iPad Model Update
Apple is also planning a refresh for its base iPad, though the timeline and ambition are both more modest. A 12th-generation iPad is expected in early 2027, featuring an updated processor — either an A18 or A19 — along with 8GB of RAM, which would bring Apple Intelligence to Apple’s most affordable tablet.
No major design changes are expected. The base iPad is being positioned to hold its ground as the entry-level option in Apple’s tablet family, accessible to students and first-time buyers who don’t need OLED panels or Pro-level chips. The addition of Apple Intelligence support via a RAM upgrade does give it a meaningful functional improvement, even if the hardware itself looks largely the same.
Gurman also noted that Apple is planning an “eventual” OLED upgrade for the iPad Air, though that’s expected sometime next year rather than this fall. iPad Pro and Apple Pencil updates could reportedly arrive in spring. Taken together, Apple appears to be running a deliberate, staged OLED rollout across its entire tablet lineup — starting with the mini this October and working outward from there. The question for buyers is whether $699 feels like the right moment to get in, or whether waiting for the iPad Air’s OLED upgrade makes more sense depending on the use case.
FAQ
What major display change is expected in the 2026 iPad mini?
The 2026 iPad mini is expected to feature an OLED display for the first time, replacing the current Liquid Retina LCD panel. The OLED upgrade brings deeper blacks, higher contrast, better HDR performance, and potentially improved battery efficiency.
Will the iPad mini’s refresh rate increase with the OLED upgrade?
No. Despite switching to OLED, the iPad mini is rumored to retain its 60Hz refresh rate. It will not feature ProMotion, Apple’s adaptive technology that enables up to 120Hz on the iPad Pro.
How will the processor change in the new iPad mini?
The new iPad mini will upgrade from the current A17 Pro chip to the more powerful A19 Pro chip, delivering faster performance particularly in demanding tasks like video editing, gaming, and AI-related features.
When and at what price is the new iPad mini expected to launch?
The new iPad mini is expected to launch as early as October 2026, with a starting price of $699 — $100 more than the current model’s price.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

A Taiwan court has handed down one of the country’s most significant Taiwan crypto fraud sentencing verdicts in recent memory, convicting a man who turned a registered crypto exchange into a front for organized crime. Shih, the ringleader behind the BitShine platform, received a 22-year prison term from the Shilin District Court after prosecutors proved he orchestrated a fraud and money laundering operation that left over 1,500 people financially devastated.
Key takeaways
Taiwan’s Shilin District Court sentenced BitShine ringleader Shih to 22 years in prison for fraud, money laundering, and illegally providing virtual asset services.
Prosecutors identified 1,539 victims who collectively lost more than NT$1.27 billion (approximately $39 million).
Between January 2024 and April 2025, the criminal group laundered more than NT$2.3 billion ($71 million), converting funds into USDT before moving them overseas.
BitShine was previously registered with Taiwan’s Financial Supervisory Commission (FSC), giving the operation a veneer of legitimacy.
The ruling arrives weeks after Taiwan passed a new law requiring all virtual asset service providers to obtain FSC approval before operating.
Court Sentences BitShine Ringleader to 22 Years
The Shilin District Court convicted Shih on three counts: illegally operating virtual asset services, orchestrating fraud, and money laundering. Prosecutors had pushed for a stiffer 25-year sentence after indicting Shih and 13 other suspects in August 2025, but the court settled on 22 years — still a remarkable term that signals how seriously Taiwan’s judiciary is treating crypto-enabled financial crime.
According to reports from the semi-official Central News Agency, prosecutors identified 1,539 victims whose combined losses exceeded NT$1.27 billion, roughly $39 million. The scale alone sets this case apart from most domestic financial fraud prosecutions.
What makes the conviction particularly striking is how Shih exploited institutional trust. BitShine was not some anonymous offshore shell — it was once a registered entity with Taiwan’s Financial Supervisory Commission. That legitimacy became the operation’s most effective weapon, reassuring victims that they were dealing with a properly supervised business while concealing the criminal machinery running beneath it.
Criminal Operations and Money Laundering Methods
The fraud did not operate in isolation. Shih’s group forged ties with fraud syndicates and affiliates connected to the Thento Union, identified by prosecutors as one of Taiwan’s three major organized crime groups. The collaboration allowed the operation to reach a scale far beyond what a standalone exchange scam could achieve.
USDT as a laundering tool
Victims’ cash was funneled into purchases of USDT — Tether’s dollar-pegged stablecoin — before being transferred overseas. The choice of a stablecoin is telling: USDT offers price stability that volatile assets like Bitcoin do not, making it a preferred vehicle when the goal is moving value across borders quickly and quietly rather than speculating. Between January 2024 and April 2025, prosecutors estimated the gang laundered more than NT$2.3 billion ($71 million) through this pipeline.
KYC procedures turned inside out
Perhaps the most cynical element of the scheme involved the exchange’s compliance infrastructure. Shih hired legitimate compliance officers — people who had no knowledge of the fraud — to build genuine know-your-customer procedures for the platform. Once those procedures were in place, intermediaries coached fraud ring members on exactly how to answer KYC verification questions, ensuring that victims could complete onboarding without triggering any red flags.
This detail matters beyond the courtroom. It illustrates how bad actors can weaponize regulatory compliance itself: building KYC systems not to screen out criminals, but to help criminals screen in victims. For regulators designing oversight frameworks, the lesson is uncomfortable — registration and procedural compliance are necessary but not sufficient safeguards.
Taiwan’s New Crypto Regulatory Framework
The sentencing lands at a pivotal moment for Taiwan’s crypto sector. Earlier this month, Taiwan’s Legislative Yuan passed the Virtual Asset Service Act, replacing the country’s previous anti-money laundering registration system with a full licensing regime. Under the new law, virtual asset service providers must obtain approval from the FSC before they can operate — a meaningful shift from the lighter-touch registration that BitShine once exploited.
The legislation goes further still. It introduces rules on cybersecurity, client asset segregation, internal controls, financial reporting, and asset listing reviews. Stablecoin issuers face the additional requirement of approval from both the FSC and Taiwan’s central bank, along with fully backed reserves held in trust, regular audits, and public disclosures.
Critically, the law now attaches criminal penalties to unlicensed operations and market abuse. Running illegal virtual asset services or issuing stablecoins without authorization can result in up to seven years in prison and fines reaching NT$100 million. Fraud and market manipulation offenses carry three to ten years and fines of up to NT$200 million. Existing firms that completed anti-money laundering registration before the law takes effect will have 12 months to apply for regulatory approval and up to 21 months to secure a full license.
The timing is not coincidental. Taiwan’s legislature passed this framework knowing that cases like BitShine had exposed the gaps in the previous registration-only system. A fraudster smart enough to register with the FSC while running an organized crime operation is exactly the kind of actor that a licensing regime with ongoing supervision — rather than a one-time registration — is designed to catch and deter.
Whether the new law will close those gaps in practice remains the real question. The BitShine case shows that determined operators with organized crime backing can build sophisticated compliance facades. The FSC will need robust ongoing monitoring, not just a tighter application process, to ensure that Taiwan’s reformed crypto sector does not become a new generation of legitimately licensed but criminally operated platforms.
FAQ
Who was sentenced in the BitShine crypto fraud case?
Shih, the ringleader behind the BitShine crypto exchange, was sentenced to 22 years in prison by Taiwan’s Shilin District Court for fraud, money laundering, and illegally providing virtual asset services.
How much money was involved in the BitShine fraud?
Prosecutors identified 1,539 victims who together lost more than NT$1.27 billion, approximately $39 million. The broader laundering operation processed more than NT$2.3 billion ($71 million) between January 2024 and April 2025.
What illegal activities was Shih convicted of by the court?
The Shilin District Court convicted Shih of illegally providing virtual asset services, orchestrating fraud, and money laundering. He had collaborated with fraud rings linked to the Thento Union and manipulated KYC procedures to facilitate the scheme.
What new regulations has Taiwan introduced regarding the crypto industry?
Taiwan recently passed the Virtual Asset Service Act, requiring all virtual asset service providers to obtain approval from the Financial Supervisory Commission before operating. The law also introduces stricter rules on cybersecurity, client asset segregation, internal controls, and financial reporting, along with criminal penalties for unlicensed operations and market abuse.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Identifying the genes that drive cancer has always been one of biology’s hardest problems. Now, a newly published framework called RegNetAgents is applying multi-agent artificial intelligence to that challenge — automating a process that once required painstaking manual curation across incompatible datasets. For researchers working at the intersection of cancer genomics AI and computational biology, the approach represents a meaningful shift in how regulatory candidates get identified and ranked.
Key takeaways
RegNetAgents is an AI-based multi-agent framework that identifies regulatory gene candidates across heterogeneous cancer networks, integrating both bulk tumor (TCGA) and single-cell (GREmLN) data.
The framework was applied to 11 breast cancer and 12 colorectal cancer focal genes, producing candidates with statistically significant enrichment for OncoKB-annotated cancer genes (all p <0.0001).
Enrichment scores reached Stouffer Z = 6.69 (TCGA, breast cancer) and Z = 7.06 (GREmLN, colorectal cancer), with no enrichment detected in housekeeping or non-driver control gene sets.
The system is implemented as a LangGraph DAG workflow accessible via a unified Python API and MCP client — functioning as a downstream layer over precomputed networks, not a network inference engine.
An extended module assesses oncogenic potential, druggability, clinical relevance, and network vulnerability to support hypothesis generation.
Why cross-network analysis changes the picture
Cancer genomics research has long struggled with a fragmentation problem. Bulk tumor sequencing data — drawn from large initiatives like TCGA — captures population-level signals across thousands of patients, but loses the cellular resolution that single-cell sequencing provides. Meanwhile, single-cell regulatory networks, like those assembled in the GREmLN project, offer granular gene-level detail that bulk data simply cannot replicate. Historically, researchers had to treat these two worlds separately.
RegNetAgents bridges that gap directly. By integrating TCGA-derived bulk tumor gene regulatory networks with GREmLN’s large-scale single-cell regulatory networks, the framework enables a unified analytical pass over both data types simultaneously. For a given focal gene of interest, the system classifies regulatory candidates drawn from each network, then ranks them by evidence consistency — flagging whether a candidate appears in both networks, in TCGA only, or in GREmLN only. That cross-network ranking is where much of the interpretive power comes from.
The scope of the initial analysis covered eleven breast cancer focal genes and twelve colorectal cancer focal genes, providing a concrete testbed across two of the most studied cancer types.
What RegNetAgents actually does
Classification, filtering, and mode-of-action assignment
At its core, the framework executes three interconnected functions for each focal gene. First, it performs dual-network classification — categorizing regulatory relationships as they appear across TCGA and GREmLN. Second, it filters candidates through OncoKB annotations, one of the most authoritative curated databases of cancer gene significance, to distinguish likely cancer-relevant regulators from background noise. Third, it assigns a mode-of-action to each tumor-derived regulatory relationship, specifying whether a candidate behaves as an activator or repressor in that context.
Together, these steps convert raw network topology into interpreted biological meaning — something that previously demanded substantial expert time.
A multi-agent LangGraph workflow under the hood
The technical architecture behind RegNetAgents is built on a LangGraph DAG (directed acyclic graph) workflow, a multi-agent design pattern that orchestrates specialized AI agents through a structured, query-driven pipeline. The system is accessible through a unified Python API and a Model Context Protocol (MCP) client, making it practical to deploy within existing computational biology environments.
Crucially, RegNetAgents is not a network inference tool. It operates as a downstream analytical layer over precomputed regulatory networks, meaning it interprets and interrogates existing network data rather than building new networks from raw expression data. That distinction matters: it keeps the system focused, computationally tractable, and interpretable — while placing the quality of upstream network construction outside its direct scope.
Performance: strong enrichment signals, clean controls
The statistical results from the breast and colorectal cancer analyses are hard to dismiss. Across TCGA-derived candidates, enrichment for OncoKB-annotated cancer genes reached a Stouffer Z score of 6.69 for breast cancer (BRCA) and 6.95 for colorectal cancer (COAD). GREmLN-derived candidates showed comparable strength: Z = 5.51 for BRCA and Z = 7.06 for COAD, with all results carrying p-values below 0.0001.
What makes these numbers more convincing is the control behavior. When the same enrichment analysis was run against housekeeping genes and non-driver control gene sets, no significant enrichment appeared. That specificity — signal in the cancer gene sets, silence in the controls — suggests the framework is not simply recovering broad biological noise but identifying candidates with genuine oncological relevance.
An extended evaluation layer for deeper insight
Beyond candidate identification, an extended module within RegNetAgents structures a deeper assessment of each shortlisted gene. This layer evaluates oncogenic potential, druggability, clinical relevance, and network vulnerability — four dimensions that collectively determine whether a regulatory candidate has real translational value. A gene might be strongly enriched in cancer networks but offer no viable therapeutic target; this module flags that distinction early.
The combination of identification and structured evaluation means the framework can carry a research question from raw network query all the way to a prioritized list of biologically interpretable hypotheses — what the authors describe as end-to-end interpretation.
Where this fits in the broader research toolbox
The arrival of RegNetAgents reflects a wider trend in computational oncology: moving from tools that generate data toward tools that interpret it. The sheer volume of regulatory network data available from TCGA, GREmLN, and comparable resources has outpaced manual analysis capacity. Multi-agent AI frameworks designed to run structured, reproducible queries across those networks address a real bottleneck.
By building the system around OncoKB cancer gene filtering, the framework also aligns candidate output with established clinical annotation standards — a practical consideration for researchers who need their computational findings to connect with existing biological knowledge.
The work was authored by Jose Bird as a PhD-level contribution and published in July 2026. Whether the framework extends cleanly to cancer types beyond breast and colorectal cancer remains an open question — one that will likely define the next phase of testing for this approach.
FAQ
What is RegNetAgents?
RegNetAgents is an AI-based multi-agent framework designed for cross-network regulatory candidate identification in cancer genomics. It integrates bulk tumor regulatory networks (from TCGA) with single-cell regulatory networks (from GREmLN) to identify and rank gene regulatory candidates relevant to cancer biology.
Which data sources does RegNetAgents integrate?
The framework integrates bulk tumor gene regulatory networks derived from the TCGA project with large-scale single-cell regulatory networks from the GREmLN project, enabling a unified analysis across both data modalities.
How does RegNetAgents evaluate candidate genes?
For each focal gene, it performs dual-network classification, filters candidates using OncoKB cancer gene annotations, and assigns a mode-of-action to tumor-derived regulatory relationships. Candidates are then ranked by evidence consistency across networks — whether they appear in both TCGA and GREmLN, or only in one.
What additional analyses does RegNetAgents provide?
An extended module evaluates each candidate’s oncogenic potential, druggability, clinical relevance, and network vulnerability, supporting comprehensive biological interpretation and hypothesis generation from identification through to translational assessment.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Three years after one of the most dramatic collapses in crypto history, FTX’s bankruptcy estate is writing checks again — and this time, the figure is roughly $900 million. The fifth wave of FTX creditor payouts is set to begin on July 31, 2026, marking another milestone in what has become a slow but steadily progressing effort to return billions to the exchange’s former users and claimants.
Key takeaways
FTX will distribute roughly $900 million to creditors starting July 31, 2026, in its fifth distribution since bankruptcy repayments began in 2025.
Total payouts from FTX’s bankruptcy estate have reached nearly $10 billion so far.
Payments will be processed within three business days through BitGo, Kraken, or Payoneer.
Creditors fall into two categories: Convenience class (retail and smaller creditors) and Non-Convenience class (larger or more complex claims).
In May 2026, law firm Fenwick & West agreed to pay $54 million to settle claims tied to enabling Sam Bankman-Fried’s fraud.
FTX Announces Fifth Creditor Distribution Starting July 31
The fifth distribution follows a pattern that has become increasingly familiar to FTX’s creditor base. According to a Friday announcement reported by The Block, eligible recipients across both the Convenience and Non-Convenience creditor classes are expected to receive their funds within three business days of the July 31 start date. Payments will flow through one of three designated services: BitGo, Kraken, or Payoneer — the same channels used in prior rounds.
For context, the fourth distribution processed in March 2026 was significantly larger at $2.2 billion. The $900 million figure for the fifth round is smaller, but when stacked against the cumulative total, it underscores just how far the bankruptcy estate has come. Since repayments kicked off in 2025, FTX has now distributed nearly $10 billion to creditors and other claimants — a figure that would have seemed implausible in the chaotic weeks following the exchange’s November 2022 implosion.
Progress and Scale of FTX Bankruptcy Distributions
The scale of what FTX’s bankruptcy estate has managed to return is genuinely striking for a crypto exchange collapse of this magnitude. The nearly $10 billion distributed so far reflects a methodical, court-supervised process under Chapter 11 that has outpaced the expectations of many industry observers who initially feared creditors would recover far less.
The distributions are structured around two distinct creditor classes. The Convenience class covers retail traders and smaller creditors, who make up the vast majority of FTX’s creditor base. The Non-Convenience class involves larger or more complex claims — institutional counterparties, sophisticated creditors, and parties with more intricate legal positions. Both groups are eligible for the July 31 payout.
FTX’s estate has aimed to repay retail creditors at a rate of 118% to 142% above the dollar value of their holdings at the time of the exchange’s 2022 collapse. That commitment to above-par recovery has been central to the estate’s communications with creditors throughout the process.
The “In-Kind” Debate and Lingering Criticism
Not everyone is satisfied. The estate has faced persistent criticism from some creditors who argue that cash repayment at 2022 valuations — even above par — fails to account for what those same crypto assets would be worth at current prices. Since many tokens have appreciated significantly since the collapse, paying in dollars rather than returning assets in kind means creditors effectively miss out on those gains.
This tension between nominal recovery rates and real-world asset performance is one of the more analytically interesting dimensions of the FTX case. A creditor who held Bitcoin at collapse-era prices and received a cash payout in the 118%–142% range may technically have been made more than whole on paper — but if they compare that to where Bitcoin or other tokens trade now, the picture looks very different. The estate’s legal structure under Chapter 11 has constrained its options, but the debate is unlikely to fade.
Fenwick & West Settlement Adds to the Recovery Pool
Beyond the direct creditor distributions, related legal proceedings have continued to add resources to the estate. In May 2026, Silicon Valley law firm Fenwick & West — which had served as FTX US’s principal outside counsel before the exchange’s collapse — agreed to pay $54 million to settle claims that it helped enable Sam Bankman-Fried‘s fraud. The firm had sought to distance itself from FTX’s conduct, but the settlement signals the breadth of legal accountability being pursued by the estate across multiple fronts.
Such settlements matter not just as symbolic accountability measures, but because they can meaningfully expand the pool of funds available for creditor recovery. Each successful claim against third parties that facilitated or overlooked fraud potentially adds to what the estate can return — a dynamic that keeps the FTX bankruptcy an active and consequential legal proceeding well into 2026.
With five distributions now either completed or underway and nearly $10 billion returned, the FTX bankruptcy has become one of the most consequential creditor recovery stories in financial history. But for those who held assets beyond the 2022 collapse-era valuations, the gap between what was lost and what was legally recovered may define how this chapter is ultimately remembered.
FAQ
When will FTX begin the fifth distribution to creditors?
FTX will start distributing the fifth payout of roughly $900 million to creditors beginning July 31, 2026.
How much has FTX paid to creditors so far in its bankruptcy process?
FTX’s bankruptcy estate has distributed nearly $10 billion to creditors and claimants since repayments began in 2025.
What types of creditors are receiving payments in the FTX distributions?
Creditors are divided into the Convenience class, which covers retail and smaller creditors, and the Non-Convenience class, which involves larger or more complex claims. Both are eligible for the current distribution.
What payment methods does FTX use to distribute funds to creditors?
Payments are processed within three business days via BitGo, Kraken, or Payoneer.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Most AI regulation happening around the world right now doesn’t actually reflect what ordinary citizens want. That’s the core finding of a new conjoint survey experiment by researchers Magnus Lundgren and Jonas Tallberg — and the implications go well beyond academic debate. When it comes to AI governance and regulation, there’s a striking gap between how governments and industries are approaching the problem and what the public is actually asking for.
Key takeaways
Citizens across seven countries with diverse political and economic profiles strongly support regulating AI.
The public prioritizes safety over innovation, public oversight over private self-regulation, and international coordination over national frameworks.
The preference for safety is most pronounced among people who see AI as risky, unpredictable, and personally consequential.
There is a systematic misalignment between dominant regulatory approaches and what citizens actually prefer.
AI’s Transformative Impact and the Regulatory Pressure It Creates
Artificial intelligence is reshaping economies, societies, and political systems at a pace that few regulatory frameworks were designed to handle. The scale of this transformation has forced policymakers into a set of genuinely difficult choices — not just technical ones, but deeply political ones about values and priorities.
The central tension is familiar: how much should governments prioritize enabling innovation versus ensuring safety? And who should be in charge — public institutions or the private sector? These aren’t hypothetical questions. They’re being answered right now, in real time, through legislation, voluntary codes of conduct, and international negotiations — often without a clear picture of what the public actually wants.
That gap between policymaker assumptions and citizen preferences is what Lundgren and Tallberg set out to measure.
What the Survey Found: Seven Countries, One Clear Signal
The researchers ran a conjoint survey experiment across seven countries chosen for their diverse political and economic profiles — a methodological choice designed to test whether preferences hold across different contexts, not just in one wealthy democracy. The breadth of that sample matters: it suggests the findings aren’t a quirk of one national culture or political moment.
The headline result is straightforward. Citizens strongly support regulating AI. This isn’t a marginal preference or a split verdict — the public broadly wants oversight to exist. What’s more interesting is how people want that oversight structured.
Safety first, innovation second
When asked to weigh safety against innovation, citizens generally came down on the side of safety. The dominant policy rhetoric in many countries — that regulation must avoid stifling innovation — doesn’t seem to match how the public frames the risk-reward calculation of AI governance.
This preference isn’t irrational. AI systems are increasingly embedded in hiring, healthcare, financial services, and law enforcement. For many people, the abstract promise of innovation feels less immediate than the concrete risk of an opaque algorithm making a consequential decision about their life.
Public oversight beats private self-regulation
Citizens favor public governance over private self-regulation — a finding that cuts against the model many technology companies have championed. Industry-led frameworks, voluntary commitments, and self-imposed ethics guidelines have been the dominant approach in several jurisdictions, particularly in the United States. The survey suggests this isn’t what the public wants.
International coordination over national frameworks
Perhaps the most geopolitically significant finding: citizens prefer international AI regulation over national approaches. In an era of fragmented, jurisdiction-by-jurisdiction rulemaking, this is a notable signal. It suggests the public intuitively understands that AI systems don’t stop at borders — and that governance probably shouldn’t either.
Risk Perception Drives the Safety Preference
Not everyone holds these preferences with equal intensity. The study identifies a clear pattern: the preference for safety in AI governance is strongest among those who perceive AI as risky, unpredictable, and personally consequential.
This finding is analytically important. It means the safety-first preference isn’t uniformly distributed — it’s amplified among people who feel directly exposed to AI’s effects. As AI becomes more visible in everyday decisions — credit scoring, medical diagnosis, content moderation — more people are likely to move into that high-concern category. If risk perception drives regulatory preference, then as AI becomes more pervasive, pressure for stronger and safer governance is likely to grow, not diminish.
It also raises a harder question: are current governance frameworks designed with the most-exposed populations in mind, or primarily around the interests of those building and deploying the systems?
The Misalignment Problem
Lundgren and Tallberg describe the core result as a systematic misalignment between dominant regulatory approaches and citizen preferences. Dominant approaches have tended to emphasize flexibility for innovation, industry self-governance, and national-level frameworks. Citizens, according to this research, want the opposite on all three dimensions.
That word — systematic — matters. This isn’t a one-off discrepancy on a single policy choice. It’s a consistent pattern across multiple dimensions of AI governance. The mismatch isn’t incidental; it reflects structural differences between how AI regulation has developed (largely driven by industry actors and national governments) and what democratic publics appear to want.
Whether that misalignment can be corrected — and through what mechanisms — is the real policy question the research leaves open. But it establishes something important: the legitimacy gap in AI governance is not just a perception problem. According to this evidence, it’s real.
FAQ
How do citizens generally feel about AI regulation?
Citizens strongly support regulating AI overall and tend to prioritize safety in AI governance, according to the conjoint survey experiment conducted by Magnus Lundgren and Jonas Tallberg across seven countries.
What governance approaches do citizens prefer for AI?
Citizens generally prefer public governance over private self-regulation and favor international regulation over national approaches — a preference that diverges from many current regulatory models.
Does risk perception affect citizens’ AI governance preferences?
Yes. Those who perceive AI as risky, unpredictable, and personally consequential show the strongest preference for safety-oriented regulation, suggesting that as AI becomes more embedded in daily life, demand for stricter oversight may intensify.
Is there alignment between citizen preferences and current AI regulatory approaches?
No. The research by Lundgren and Tallberg finds a systematic misalignment between dominant regulatory approaches — which tend to favor innovation flexibility and self-regulation — and the preferences citizens actually express.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Anthropic is rethinking how developers should deploy AI models — not as isolated tools, but as coordinated systems where each model plays a specific role. At the center of that shift is Claude Platform orchestration, a framework that pairs the powerful Fable 5 model with lighter, faster alternatives to get frontier-level results without paying frontier-level token costs.
Key takeaways
Claude Platform now includes four models — Fable 5, Opus 4.8, Sonnet 5, and Haiku — each with a distinct role in the AI stack.
Fable 5 is built for frontier reasoning and long-horizon agentic work, acting as a strategic advisor rather than a hands-on executor.
The advisor strategy routes heavy thinking to Fable 5 while smaller, cheaper models like Sonnet 5 and Haiku handle execution.
Developers can build custom eval suites scoped to their own tasks to decide which work should move to Fable 5.
Cost management tools include prompt caching, batch processing, and task budgets.
Anthropic’s Claude Platform Model Lineup and Roles
Anthropic’s model lineup has never been this wide. Each model in the Claude Platform serves a clearly defined purpose, and understanding those distinctions is the starting point for any serious deployment strategy.
Fable 5 sits at the top, designed specifically for frontier reasoning and long-horizon agentic work — the kind of complex, multi-step tasks that require sustained planning and high-level judgment. Below it, Opus 4.8 handles everyday complex tasks that still demand serious cognitive load. Sonnet 5 functions as the platform’s default workhorse, balancing capability with efficiency. And Haiku is optimized for speed and scale, ideal when throughput matters more than depth.
That differentiation isn’t just a product taxonomy. It reflects a deliberate design philosophy: not every task deserves the most powerful model, and throwing Fable 5 at every request would be both wasteful and unnecessary.
Innovative Model Orchestration Using Fable 5 as Advisor
The advisor strategy is the most analytically interesting part of Anthropic’s framework. Rather than using Fable 5 as the primary executor of tasks, developers are encouraged to deploy it as a strategic advisor — a high-level planner that sets direction while smaller, cheaper models do the actual work.
In practice, this means Fable 5 handles the reasoning and delegation layer, while Sonnet 5 or Haiku execute the individual steps. According to Anthropic, this pattern can match frontier-level results at a fraction of the token cost — a claim that makes the approach particularly attractive for teams managing cost at scale.
The complementary orchestrator strategy adds another layer: deciding when Fable 5 or Opus 4.8 should plan and delegate, versus when Sonnet 5 and Haiku should step in and execute. Together, the advisor and orchestrator patterns give platform engineers a concrete decision framework for multi-model pipelines rather than ad hoc model selection.
This matters beyond the efficiency argument. As AI agents take on longer, more autonomous workflows, the ability to chain models intelligently — rather than relying on a single model for everything — becomes a structural advantage. Teams that build these orchestration patterns now are effectively constructing an architectural moat that persists across future model upgrades.
Developer Tools: Custom Eval Suites and Cost Management
Knowing which model to use in theory is one thing. Knowing which model to use for your specific tasks is another. That’s where custom eval suites come in.
Brad Abrams, Product Manager at Anthropic, and Jeremy Hadfield from Anthropic’s Applied AI team walk developers through how to build evaluation suites scoped directly to their own workloads — suites designed to survive model upgrades and provide consistent guidance on when it makes sense to route work to Fable 5 versus a lighter alternative.
The practical cost management layer covers several techniques:
Prompt caching to reduce redundant token processing
Batch processing for high-volume, latency-tolerant tasks
Task budgets and effort levels to cap resource usage per workflow
These aren’t abstract optimizations. For teams operating at scale on the Claude API, the difference between an unmanaged deployment and one using prompt caching and batch processing can translate directly into significant infrastructure cost reductions.
The session is aimed squarely at heads of AI, platform engineering and architecture leaders who own model strategy and internal evals, and developers actively building agents and orchestration pipelines on the API. It’s a technical audience, and the content reflects that — this isn’t an introduction to AI, it’s a blueprint for production-grade deployment.
What makes the overall framework compelling isn’t any single feature — it’s the integrated logic. A well-designed eval suite tells you which tasks belong to which model. The advisor strategy tells Fable 5 how to set strategy without burning tokens on execution. And cost controls keep the whole system sustainable. The question facing most platform teams now is how quickly they can operationalize these patterns before the cost and complexity of uncoordinated multi-model deployments catches up with them.
FAQ
What models are included in Anthropic’s Claude Platform?
Claude Platform includes Fable 5 for frontier reasoning and long-horizon agentic work, Opus 4.8 for complex everyday tasks, Sonnet 5 as the default workhorse, and Haiku for speed and scale.
How does the model orchestration strategy work on the Claude Platform?
Fable 5 acts as a strategic advisor, setting direction and delegating tasks, while smaller and cheaper models like Sonnet 5 and Haiku handle execution. This pattern is designed to match frontier-level results at a lower token cost.
Can developers customize how tasks are assigned to models in Claude Platform?
Yes. Developers can build custom evaluation suites scoped to their own tasks to determine which work should be routed to Fable 5 and which can be handled by lighter models. These suites are also designed to remain useful across future model upgrades.
Who is the target audience for learning about Claude Platform orchestration?
The framework is aimed at heads of AI, platform engineering and architecture leaders responsible for model strategy and internal evaluations, and developers building agents and orchestration systems on the Claude API.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Coinfest Asia, The World’s Crypto Festival organized by Indonesia Crypto Network (ICN), will return to Melasti Beach, Bali, Indonesia, on 20–21 August 2026, bringing together institutions, builders, traders, founders, investors, developers, and global Web3 communities in one of Asia’s most dynamic crypto gatherings.
This year, Coinfest Asia will feature dedicated programs for different parts of the crypto ecosystem, helping attendees access relevant insights, networking opportunities, product showcases, and business connections across Asia’s fast-growing crypto and Web3 markets.
What Attendees Can Do at Coinfest Asia
The 2026 edition departs from traditional conference formats by organizing content into three intent-based tracks. This structure is intended to align attendees with specific functional areas of the industry:
Institutional Track
Focused on digital asset adoption, stablecoin integration, and tokenization. Programs include Stablecoins Readiness Workshop, Asia Go-To-Market Sessions, Closed-Door Stablecoins & Tokenization Roundtable, panel discussions, and keynote sessions.
Builders Track
Geared toward developers and startups across AI, blockchain, and digital infrastructure. Programs include Gemini AI Masterclass, “What the Hack!” Web3 Developer Course, AI Vibe Code Competition, Asia Go-To-Market Sessions, and sessions focused on product development and ecosystem growth.
Traders Track
Created for active traders and market participants looking to understand market narratives, sharpen trading strategies, and connect with trading communities. Programs include Trading Competition by TRIV, Alpha Hunting Masterclass, Yapper Masterclass, Bitcoin Crash Course, Live Degen Experience, trading-focused panels, and keynote sessions.
Through these tracks, Coinfest Asia aims to make the festival easier to navigate while keeping the experience open and connected across the wider crypto industry.
Opening Access to Asia’s Web3 Markets
A core objective of the 2026 event is providing localized insight into Asian markets. The event introduces “Asia Go-To-Market Sessions,” which provide briefings on regulatory environments, user behaviors, and growth channels in specific jurisdictions.
These sessions are organized in collaboration with regional ecosystem partners:
Japan GTM Session with WebX 2026
Malaysia GTM Session with MYBW 2026
Vietnam GTM Session with Conviction
Indonesia GTM Session with Indonesia Crypto Network
Taiwan GTM Session with FutureMode
India GTM Session with India Blockchain Week 2026
Through these sessions, attendees can better understand local user behavior, regulatory direction, community dynamics, partnership opportunities, and distribution strategies across Asia.
“Asia is not one single market. Each country has its own users, regulations, culture, and growth channels,” said Joditha Winatajaya, Head of Event at Coinfest Asia. “Through Asia Go-To-Market Sessions, we want to connect the audiences with the right local ecosystems, all in one place.”
A Foundation Built on Industry Leadership
Coinfest Asia 2026 will feature speakers from across blockchain infrastructure, exchanges, wallets, stablecoins, payments, data, AI, institutional finance, venture capital, to Web3 communities.
Confirmed speakers include Charles Hoskinson (Founder, Input Output Group), Felix Fan (CEO, Trust Wallet), Alexander Svanevik (CEO, Nansen), Nick See Tong (APAC & Singapore Lead, Base), Iñaki Moreno (Strategic Partnerships Lead, Web3 & AI, Google), William Sutanto (CEO, INDODAX), Ploy Boonyavee (Thailand/Indochina Country Manager, Tether), Gabriel Rey (CEO, TRIV Group), Calvin Kizana (CEO, Tokocrypto), Angela Ang (APAC Managing Director and Singapore President, BitGo), Tianwei Liu (CEO, StraitsX), Thomas Chou (Head of APAC, Canton Foundation), Akshat Vaidya (Co-Founder, Maelstrom) and more.
The event is also supported by leading companies across crypto, fintech, infrastructure, and digital assets, including Binance, Tokocrypto, Indodax, Triv, Duitku, ClickHouse, CockroachDB, BYDFi, Zoomex, FundedXyz, WalletConnect, GOIDR, GudangKripto, and more partners to be announced.
Since its launch, Coinfest Asia has grown into one of the world’s leading crypto gatherings, bringing together global companies, local ecosystems, builders, traders, institutions, and communities in Bali.
The 2026 edition builds on that momentum by combining industry programming with a festival environment designed for more fluid interaction. Beyond the main stages, Coinfest Asia will feature expo areas, curated business matching, networking activations, product showcases, community gatherings, and side events across the festival experience.
Tickets for Coinfest Asia 2026 are now available. Companies looking to expand into Asian crypto markets can also explore partnership and marketing opportunities through the official event channels.
About Coinfest Asia
Coinfest Asia is the world’s largest crypto festival, organized by Coinvestasi, a subsidiary of Indonesia Crypto Network (ICN). Held annually in Bali, Indonesia, the event brings together institutions, builders, and traders to connect, collaborate, and drive the future of digital assets in Asia and beyond.
Learn more about Coinfest Asia.

A crypto firm is about to put its name on one of college football’s most storied venues — and the Galaxy Digital Texas Tech deal signals something bigger than a stadium sign swap. Starting with the 2026 season, the Lubbock home of Red Raider football will carry a new identity: Galaxy Stadium, replacing the Jones AT&T Stadium name that had stood for nearly eight decades.
Key takeaways
Galaxy Digital signed a 15-year naming rights partnership with Texas Tech, effective from the 2026 football season.
Jones AT&T Stadium will be officially rebranded as Galaxy Stadium, opening its new era on September 5 against Abilene Christian.
The deal reportedly exceeds $70 million in total value, or approximately $4.7 million annually, according to Yahoo Sports’ Ross Dellenger.
Galaxy becomes Texas Tech’s official data center and digital assets partner, with branding across football and basketball programs.
Galaxy’s Helios data center campus in nearby Dickens County carries 1.6 gigawatts of approved capacity and already employs Texas Tech graduates.
Galaxy Digital’s 15-Year Partnership with Texas Tech
The agreement goes well beyond a naming rights deal. Galaxy Digital (Nasdaq: GLXY) is embedding itself into the Texas Tech athletic infrastructure at multiple levels — from student-athlete endorsements to academic pipeline development — in what the company calls the first step of a broader strategic relationship.
Stadium Renaming to Galaxy Stadium
The home of Red Raider football opens its renamed chapter on September 5, 2026, when Texas Tech hosts Abilene Christian. Texas Tech is coming off a Big 12 title and College Football Playoff appearance, giving the stadium’s new identity an immediate high-profile stage.
The Jones name honored the university’s third president, Clifford B. Jones, whose original gift helped establish the football program. Texas Tech confirmed that the Jones family’s legacy will be formally recognized, with details on a tribute to be announced separately.
“We look forward to creating many more of those moments together in Galaxy Stadium, one of the premier home-field environments in college football,” said Kirby Hocutt, Texas Tech’s Director of Athletics. The naming deal was facilitated by Texas Tech Athletics Partners through Learfield, the university’s exclusive multimedia rights holder.
Official Data Center and Digital Assets Partnership
Naming rights are only the surface layer. The partnership designates Galaxy as the official data center and digital assets partner of Texas Tech Athletics, with branding extended across Red Raider football and both men’s and women’s basketball — through in-game features, digital channels, and social media platforms.
The agreement also opens the door to exploring academic and commercial applications of artificial intelligence, workforce development initiatives, and broader economic investment across West Texas. It’s a structure that frames the deal less as a sponsorship and more as a long-term regional presence play.
Financial and Branding Aspects of the Deal
Reported Valuation of Over $70 Million
Galaxy did not disclose financial terms in its official release. However, Yahoo Sports reporter Ross Dellenger reported the 15-year agreement is worth more than $70 million, working out to roughly $4.7 million annually. For context, that puts it squarely in the upper tier of collegiate naming rights deals.
Athlete Endorsement and Marketing Collaborations
The deal includes Name, Image and Likeness (NIL) opportunities for Texas Tech student-athletes through branded activation campaigns and original content. Galaxy will work directly with Red Raider athletes on endorsement deals and marketing campaigns — a structure that reflects how modern collegiate partnerships increasingly bypass traditional advertising in favor of athlete-driven content.
Andrew Wheeler, Executive Vice President of Sports Properties at Learfield, called it “a truly modern partnership” that integrates “student-athlete storytelling, enhanced fan experiences, community, and campus-wide impact” rather than functioning as a conventional naming rights transaction.
Strategic Implications for Tech Talent and Regional Development
Helios Data Center’s Role and Local Hiring
Galaxy’s strategic rationale is anchored in geography. Its Helios data center campus, located in Dickens County roughly 60 miles east of Lubbock, carries 1.6 gigawatts of approved capacity for high-performance computing, positioning it among the largest data center developments in North America. Texas Tech graduates are already working there, and the partnership is explicitly designed to deepen that pipeline.
The company has invested billions in its West Texas buildout, a significant share of which flows through the Lubbock economy. Expanding the hiring pipeline from a major regional university isn’t just good optics — it’s a practical workforce strategy for a campus-scale data infrastructure operation that needs sustained technical talent.
Mike Novogratz’s Vision for West Texas
Galaxy founder and CEO Mike Novogratz framed the deal in explicitly regional and economic terms. “Texas Tech has a culture built on grit and loyalty, one of the strongest talent pipelines in the country,” he said. “We’re building the infrastructure that powers the code economy — and we’re doing it the right way: prioritizing hiring locally, investing in the community and being a good neighbor.”
That framing matters. It positions Galaxy not as a crypto firm buying visibility, but as a permanent infrastructure investor staking a long-term claim in West Texas. Whether the ambition matches execution over 15 years remains to be seen, but the structural logic — proximity to a major research university, an already operational gigawatt-scale data center, and a region hungry for economic investment — is coherent.
Context of Crypto and AI Infrastructure in Collegiate Sports
The Galaxy deal doesn’t exist in isolation. It’s part of a visible acceleration of crypto and AI infrastructure firms using sports partnerships to build brand recognition and talent pipelines simultaneously.
Ripple’s University of Kansas Sponsorship
Just days before the Galaxy announcement, Ripple became an official sponsor of the University of Kansas — the alma mater of Ripple CEO Brad Garlinghouse. The deal made XRP the first cryptocurrency to appear on the jerseys of a major collegiate athletics team. Ripple also committed to financial and technology education for student-athletes and expanding its own recruiting pipeline from the university.
IREN’s Golden State Warriors Partnership
On the professional side, AI cloud provider IREN — a former pure-play bitcoin mining company — signed a jersey-patch agreement with the Golden State Warriors. Sportico reported the deal is worth more than $50 million annually, making it the largest sponsorship deal in North American sports by that measure.
Taken together, these deals suggest a structural shift, not a coincidence. Crypto and AI infrastructure companies are discovering that sports partnerships offer something their sector has historically lacked: sustained mainstream visibility, a direct line to technical talent through university recruiting pipelines, and community credibility in regions where they’re building physical infrastructure. For Galaxy, putting its name on a 60,000-seat football stadium in West Texas is as much a local roots play as it is a branding move — and that dual purpose is what distinguishes it from a simple marketing spend.
FAQ
What is the duration and scope of Galaxy Digital’s partnership with Texas Tech?
Galaxy Digital signed a 15-year agreement with Texas Tech covering stadium naming rights and designation as the university’s official data center and digital assets partner, with branding across football and basketball programs and NIL opportunities for student-athletes.
When will Jones AT&T Stadium be renamed Galaxy Stadium?
The stadium will officially become Galaxy Stadium beginning with the 2026 football season, with the first game under the new name scheduled for September 5, 2026, against Abilene Christian.
What are some of the branding and marketing aspects included in the partnership?
Galaxy will collaborate with Texas Tech athletes on endorsement deals and marketing campaigns through NIL activations, and will receive branding across football and men’s and women’s basketball programs both at games and through Texas Tech Athletics’ digital and social media platforms.
How does Galaxy Digital plan to leverage local talent through this partnership?
Galaxy expects to expand its hiring pipeline from Texas Tech graduates to its Helios data center campus in Dickens County, which already employs Texas Tech alumni and carries 1.6 gigawatts of approved capacity for high-performance computing.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

A free AI model from Beijing just rattled crypto markets, and the timing could hardly have been more pointed. When Moonshot AI dropped Kimi K3 on Thursday, it didn’t just shake up Silicon Valley’s AI pecking order — it sent Bitcoin and ether sliding, revived memories of the DeepSeek shock, and forced a hard question onto the table: if frontier AI capability is now free, what exactly are all those billions in infrastructure bets actually buying?
Key takeaways
Moonshot AI released Kimi K3, a 2.8 trillion-parameter open-weight model, with full public weights scheduled for July 27 — free to download and run.
Kimi K3 scored 1,679 on Arena’s Frontend Code leaderboard, beating Anthropic’s Claude Fable 5 (1,631) and OpenAI’s GPT-5.6 (1,618).
Bitcoin and major cryptocurrencies fell after the release, with traders calling it a “Kimi moment” echoing the DeepSeek shock.
Chinese AI rivals Z.ai and MiniMax saw stocks drop 27% and 16% respectively following the announcement.
Bitcoin is increasingly behaving as a leveraged bet on the AI capital cycle, moving with semiconductor and AI infrastructure sentiment rather than on-chain events.
Moonshot AI launches Kimi K3 as the world’s largest open-source AI model
Moonshot AI released Kimi K3 on Thursday, and within hours the market had a name for what followed: a “Kimi moment.” The reaction drew immediate comparisons to DeepSeek’s January 2025 debut, which wiped roughly $600 billion from Nvidia’s market cap in a single session. This time, the shockwave hit both AI stocks and crypto simultaneously.
The model itself is a genuine engineering milestone. At 2.8 trillion parameters — roughly 75% larger than DeepSeek’s V4 Pro — it is, according to Moonshot, the largest open-source AI model ever built. It carries a one-million-token context window, native visual understanding, and an always-on reasoning mode the company calls “thinking mode.”
Architecture built for efficiency, not just scale
What makes the scale sustainable is the underlying design. Kimi K3 runs on a mixture-of-experts architecture, activating only 16 specialist modules out of 896 for any given task. That selectivity is what keeps a model of this size cheap to run. Moonshot says architectural changes — including internally developed techniques called Kimi Delta Attention and Attention Residuals — deliver roughly 2.5 times the scaling efficiency of its predecessor.
The model is also built for developer integration. It is compatible with the OpenAI SDK, meaning engineers already working on Anthropic or OpenAI toolchains can slot it in with minimal friction. On the API, it is priced at $3 per million input tokens and $15 per million output tokens — mid-tier pricing for what the company claims is a top-tier product.
Benchmark results: coding crown, not overall throne
On Arena’s Frontend Code leaderboard, Kimi K3 scored 1,679, ahead of Anthropic’s Claude Fable 5 at 1,631 and OpenAI’s GPT-5.6 at 1,618. That put it first overall, ranking top in six of seven categories. Moonshot’s own previous model had sat at number 18 — a 17-place jump in a single release.
The caveat matters, though. On broader general knowledge and task benchmarks, Kimi K3 still trails the top Claude and OpenAI configurations. This is a dominant win in a specific, high-value domain. Not a sweep across the board. Bank of America analysts, in a note led by Alex Liu, framed it pointedly: “K3 raises the capability ceiling for China AI models, shifting the burden of proof to other independent AI labs.”
Moonshot also showcased a proof-of-concept that hints at longer ambitions. Over 48 hours of continuous autonomous operation, K3 independently designed a functional chip — reading documentation, making architectural decisions, running verification loops — using open-source electronic design automation tools. The result was a 4-square-millimeter chip design achieving timing convergence at 100 MHz. It is not a production chip. It is a signal of where Moonshot thinks the next competitive edge lies.
Public release and open accessibility
The full model weights are scheduled for public release on July 27. Anyone will be able to download and run Kimi K3 on their own hardware, at no cost. That open-weight commitment — combined with the coding benchmark win — is the core of why markets reacted the way they did.
Market reactions and impact on AI and cryptocurrency sectors
Bitcoin, ether, and effectively every major cryptocurrency fell on Friday after the Kimi K3 release. The sell-off was not driven by anything happening on-chain. It was a macro sentiment trade rooted in what a free Chinese AI model implies for the economics of AI infrastructure.
Cryptocurrency price declines following the Kimi K3 release
The crypto market’s reaction to the Kimi AI model impact ran directly through AI infrastructure logic. Earlier in the same week, Bitcoin had risen 4% on the day South Korea’s Kospi jumped 8% and SK Hynix priced $26.5 billion of American depositary shares. The same AI compute trade that lifted prices one Friday knocked them lower the next. The symmetry is hard to ignore.
Patrick Moorhead, CEO of Moor Insights and Strategy, called the market’s response “an over-reaction shockingly similar to the DeepSeek panic,” arguing on X that models like Kimi K3 will “accelerate and grow the inference market faster than without.” His framing — that the overall reaction was driven partly by Washington politics around Chinese AI adoption — adds a layer of complexity the price charts don’t fully capture.
Chinese AI rivals take the biggest hit
Moonshot’s domestic competitors absorbed the sharpest blow. Z.ai, which had released a new model to significant fanfare in June, fell approximately 27%. MiniMax Group dropped around 16%. Alibaba, whose Qwen open-source narrative was already under pressure, slid 4% — even as it had been buoyed earlier in the week by news of a partnership with Apple in China.
Liu’s note from Bank of America put it directly: Alibaba’s position as “open-source leader” now faces a meaningful test from a rival that has just set a new scale record.
Challenges to AI infrastructure investment assumptions
The deeper disruption here is not about one benchmark. It is about the assumption that has been quietly underwriting hundreds of billions of dollars in AI infrastructure spending: that frontier capability stays scarce, expensive, and American.
A free model at the top of a major coding leaderboard is a direct counterargument. Perplexity CEO Aravind Srinivas captured a related shift last week when he told CNBC that “the model alone is no longer the product” — it is the orchestration harness around it. If models themselves become commodities, the entire capital stack built on model scarcity looks different. Lu Zhang of Fusion Fund noted that most of the developers who would actually adopt Kimi K3 come from the startup ecosystem rather than large corporates, and that high-powered AI models still require significant technical expertise to deploy in production — a real constraint on how fast the shift plays out.
Bitcoin’s evolving market dynamics amid AI capital cycle shifts
What the Kimi K3 episode clarifies is something that has been building for months: Bitcoin is no longer simply a crypto asset trading on crypto-specific catalysts.
Bitcoin price movement increasingly tied to AI infrastructure sentiment
Bitcoin has spent the past week taking direction from semiconductor and AI infrastructure sentiment rather than anything happening on-chain. The pattern is consistent enough that it is hard to dismiss as coincidence. A chip listing in Seoul moves Bitcoin one direction; a model release in Beijing moves it the other. The on-chain world — hash rates, exchange flows, ETF inflows — is not driving price at the margin right now.
Bitcoin miners’ pivot toward AI data centers
There is a concrete reason for this exposure. Bitcoin miners have spent roughly two years repositioning themselves as AI data center landlords, signing long-term leases with model developers on the premise that demand for training and inference compute would keep rising. Several public Bitcoin companies have staked significant portions of their forward revenue on this thesis.
That thesis prices in scarcity. If a free, open-weight model running efficiently on less hardware can sit at the top of a coding leaderboard, the tenants those miners are counting on have less structural reason to sign long-term compute contracts. The floor under the miner-to-AI pivot becomes less certain with every open-weight release that erodes the premium on proprietary compute access.
Bitcoin as a leveraged expression of the AI capital cycle
The comparison to January 2025 is instructive, but the situation is not identical. When DeepSeek dropped eighteen months ago, Bitcoin sold off as a risk asset in a broad risk-off session. What is different now is the nature of the exposure. In July 2026, as reported by CoinDesk, Bitcoin is trading as a leveraged expression of the AI capital cycle itself — up on a Korean chip listing one week, down on a Chinese model release the next.
After DeepSeek, Nvidia recovered. Bitcoin recovered. Capital expenditure kept climbing. The market’s memory of that episode might cushion the immediate reaction to Kimi K3. But the structural question is sharper now because Bitcoin’s positioning inside the AI capital cycle is more explicit than it was eighteen months ago. The full model weights land on July 27. At that point, the market stops trading on a benchmark score and starts trading on whether the capability holds up in production — and what that means for the infrastructure bet that has quietly become central to Bitcoin’s identity as an asset.
FAQ
What is Kimi K3 and who developed it?
Kimi K3 is an open-weight coding AI model developed by Beijing-based Moonshot AI. It features 2.8 trillion parameters, a one-million-token context window, and a mixture-of-experts architecture that keeps running costs low despite its scale.
How does Kimi K3 perform compared to other AI models?
On Arena’s Frontend Code leaderboard, Kimi K3 scored 1,679, outperforming Anthropic’s Claude Fable 5 (1,631) and OpenAI’s GPT-5.6 (1,618). However, on broader general knowledge benchmarks, it still trails the top configurations from both Anthropic and OpenAI.
How did the release of Kimi K3 affect cryptocurrency markets?
Following Kimi K3’s release, Bitcoin and other major cryptocurrencies fell as traders drew parallels to the DeepSeek shock. The sell-off reflected concerns about the AI infrastructure thesis that has increasingly underpinned Bitcoin miner valuations, rather than any crypto-specific on-chain development.
Why does Kimi K3 challenge current AI infrastructure investment assumptions?
Because Kimi K3 is free and open-weight, it directly undermines the assumption that frontier AI capability will remain scarce, expensive, and controlled by U.S. companies. That assumption has justified massive infrastructure spending — and a free, high-performing model weakens the economic case for locked-in compute contracts.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Wenn eine Krypto-Plattform ausfällt und Nutzer von ihren eigenen Konten ausschließt, kommt die rechtliche Maschinerie selten schnell genug, um noch etwas zu bewirken. Im Fall von Knaken Cryptohandel BV, einer in Rotterdam ansässigen Kryptobörse, geschah jedoch das Gegenteil – und was sie vorfand, war ein düsteres Bild. Ein niederländisches Gericht hat das Unternehmen und seine angeschlossene Stiftung für bankrott erklärt, nachdem 7 Millionen Euro an Kundengeldern verschwunden waren. Dadurch bleibt eine unbekannte Zahl von Nutzern mit nur wenig Informationen darüber, ob sie ihr Geld jemals wiedersehen werden.

A venture capitalist has found a blunt workaround to one of Silicon Valley’s most quietly contentious habits. On Zoom, Jeremy Levine no longer simply logs in as himself — his display name now reads “Jeremy Levine I do not consent to transcribing or recording.” It’s part protest, part practical shield, and it says something uncomfortable about where Zoom AI transcription and always-on recording have taken us.
Key takeaways
Venture capitalist Jeremy Levine changed his Zoom display name to formally state his non-consent to recording or transcription.
VC Eric Bahn now automatically assumes every meeting with a founder will be recorded, even without seeing a device.
A founder uses the Granola app to record first dates, then feeds transcripts to Claude to analyze her own conversational behavior.
Always-on recording is described as a legal minefield, raising unresolved questions about consent.
The sheer volume of auto-generated transcripts is creating a new problem: recordings that no one has time to revisit.
The rise of always-on AI recording
A new Wall Street Journal report captures what many professionals have already started to feel in their bones: the assumption of being recorded has quietly become the default in modern meetings. A growing ecosystem of AI note-taking apps and wearable devices has made continuous transcription not just possible but normal — expected, even.
TechCrunch, which has covered and ranked multiple tools in this space, reported on the trend July 17, 2026. The shift isn’t subtle. VC Eric Bahn told the Wall Street Journal that he now automatically assumes his meetings with founders will be recorded — and that assumption kicks in before anyone even slides a phone across the conference table. The recording is already happening. He just knows it.
That level of ambient awareness marks a real turning point. When a senior investor treats recording as a background constant rather than an active choice by the other party, it signals how thoroughly these tools have normalized surveillance in professional settings.
When AI transcription moves beyond the office
What makes the current moment genuinely strange is how far outside work this behavior has traveled. According to the Wall Street Journal piece, one founder disclosed that she records most of her first dates using the Granola app. After each date, she feeds the transcript to Claude — an AI tool — to assess whether she could have been more “engaging or empathetic,” and to figure out who did most of the talking.
That’s not a productivity workflow. That’s using AI transcription as a personal performance coach for romance.
It’s a vivid example of how tools built for conference rooms are bleeding into the most intimate corners of daily life. And it raises an obvious question that nobody seems to have a clean answer for: does the other person know?
The social and legal fallout nobody wants to talk about
Levine’s Zoom name stunt reads as frustration turned into a public statement. He has called the always-on recording trend “socially unacceptable behavior” that can kill spontaneous conversation entirely. When people know — or simply suspect — that their words are being logged, the texture of interaction changes. The offhand remark, the candid admission, the kind of thinking-out-loud that moves a conversation forward: all of it gets filtered through a new layer of self-censorship.
Beyond the social friction, experts cited in the Wall Street Journal piece describe the legal terrain as a minefield. Recording consent laws vary widely, and the casual deployment of AI transcription apps — often without explicit notification to all parties — sits in genuinely murky legal territory. The gap between what’s technically easy and what’s legally permissible has rarely been this wide.
There’s a strategic dimension here worth examining. Companies building AI note-taking products have strong incentives to make recording as frictionless as possible. Every added consent prompt is a moment of friction that reduces usage. But that same frictionlessness is precisely what creates the legal and social exposure. The easier it becomes to record without thinking, the harder it becomes to argue the recording was meaningfully consented to.
The data overload problem no one anticipated
Even setting aside consent, there’s a practical absurdity beginning to surface. If every meeting, casual conversation, and first date gets auto-transcribed and summarized, who is actually reading any of it?
The honest answer, in most cases, is nobody. Recordings pile up. Summaries go unopened. The initial appeal of having everything captured collides with the reality that human attention is finite. What was supposed to solve information loss ends up creating a different kind of problem: an audio archive of daily life that nobody has the time or energy to process.
This points to a tension at the core of the always-on recording movement. The value proposition assumes that capturing everything preserves value. But captured data without retrieval isn’t memory — it’s just storage. And storage, at scale, becomes its own burden.
Levine’s Zoom name hack won’t stop the trend. But it does put the consent question squarely on the screen, literally, for anyone who joins his calls. The real test isn’t whether individuals can opt out through clever display-name workarounds — it’s whether the companies building these tools will build consent in by default, before regulators force their hand.
FAQ
Why did Jeremy Levine change his Zoom name?
Levine changed his Zoom display name to “Jeremy Levine I do not consent to transcribing or recording” to formally signal his refusal to be recorded or transcribed during video meetings, in response to the widespread use of AI transcription tools.
Are AI transcription apps commonly used in meetings now?
Yes. According to reporting by the Wall Street Journal and TechCrunch, always-on recording has become increasingly ubiquitous, driven by a growing range of AI note-taking apps and devices used across professional and personal settings.
What legal issues does continuous recording raise?
Continuous recording without explicit consent from all parties raises serious legal questions. Sources cited in the Wall Street Journal describe the practice as a legal minefield, though the specific laws involved vary depending on jurisdiction.
How do some individuals use AI transcription in personal contexts?
One founder, as reported by the Wall Street Journal, records most of her first dates using the Granola app, then feeds the resulting transcripts to Claude to evaluate how engaging or empathetic she was and to see how much of the conversation she led.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Something quiet happens when a language model gets fine-tuned to be helpful and safe — it starts to sound like everyone else. That homogenization effect, known as mode collapse, has long been treated as a byproduct of imperfect training algorithms. But new research challenges that assumption at its root, tracing the problem not to the algorithm, but to the data itself — and specifically to a deeply human cognitive quirk baked into every preference label.
Key takeaways
Post-training alignment reduces diversity in large language models, producing mode collapse — a tendency to generate repetitive, predictable outputs.
The root cause is typicality bias in preference data: human annotators consistently favor familiar-sounding text, which shapes model behavior at scale.
Typicality bias is grounded in cognitive psychology and has now been formalized theoretically and verified empirically on preference datasets.
Verbalized Sampling (VS) is a training-free prompting method that counteracts mode collapse by asking the model to generate multiple responses alongside probability estimates.
In creative writing tasks, VS increases output diversity by 1.6 to 2.1 times compared to direct prompting, with gains across dialogue, open-ended QA, and synthetic data generation.
Mode Collapse Driven by Post-Training Alignment and Typicality Bias
Mode collapse mitigation has become one of the more pressing challenges in modern AI development, precisely because the problem is so easy to miss. A fine-tuned model still answers questions. It still writes poems. It just writes the same kind of poem, over and over, in slightly different words.
Impact of Post-Training Alignment on LLM Diversity
Post-training alignment — the process by which a base language model is shaped through human feedback to be more helpful, harmless, and honest — consistently reduces the generative diversity of large language models. The result is a narrowing of the output space: models converge on a smaller set of “acceptable” responses, trimming the tails of their distribution in ways that sacrifice originality and variety.
This is not a minor stylistic concern. For applications like synthetic data generation, creative writing, or dialogue simulation, diversity is a functional requirement. A model that collapses toward the mean produces training data that reinforces the same biases, dialogue that feels scripted, and creative output that feels derivative.
Role of Typicality Bias in Preference Data
Typicality bias is the mechanism at the heart of the problem. When human annotators evaluate model outputs and label which responses they prefer, they systematically favor text that feels familiar — responses that match their intuitive sense of what a “typical” good answer looks like. This preference for the prototypical over the novel is not a flaw unique to AI labelers; it reflects well-established findings from cognitive psychology about how humans categorize and evaluate information.
The consequence, at scale, is significant. Preference datasets built from thousands of such annotations encode a structural bias against unusual but valid responses. Models trained on this data learn, implicitly, that unfamiliar outputs are less desirable — even when those outputs are correct, creative, or meaningfully different from the median. Typicality bias in AI systems, in other words, is an inheritance from human cognition passed through the training pipeline.
Theoretical and Empirical Analysis of Typicality Bias
The research behind these findings does not stop at identifying the problem. The authors formalize typicality bias theoretically, constructing a rigorous framework that explains how annotator preferences distort the learned distribution of aligned models. They then verify this effect empirically, testing it against real preference datasets to confirm that the bias is not incidental but pervasive and central to the mode collapse phenomenon.
This dual approach — theoretical formalization followed by empirical validation — matters because it shifts mode collapse from a vague observation into a tractable, well-defined problem. It also opens the door to principled solutions, rather than ad hoc engineering fixes. By understanding why alignment narrows output diversity, researchers can design methods that address the cause rather than mask the symptom.
Verbalized Sampling: A Training-Free Approach to Mitigate Mode Collapse
The proposed remedy is called Verbalized Sampling, and its core insight is elegant in its simplicity. Rather than retraining the model or modifying the alignment pipeline — both expensive and technically demanding — Verbalized Sampling works entirely at inference time, through a change in how the model is prompted.
Mechanics of Verbalized Sampling Prompting
Instead of asking a model to produce a single response, Verbalized Sampling instructs the model to generate a set of candidate responses and assign explicit probability estimates to each. A prompt might read: “Generate 5 jokes about coffee and their corresponding probabilities.” By forcing the model to reason over a distribution of possible outputs rather than committing to one, VS bypasses the mode-collapsing tendency that alignment has instilled. The model’s pre-trained generative diversity, which was suppressed but not erased by fine-tuning, gets reactivated through this probabilistic framing.
The practical appeal here is real. No retraining. No new datasets. No changes to the model architecture. The method is applicable to any aligned language model, and it introduces no additional infrastructure burden.
Performance Improvements Across Multiple Tasks
Experiments testing Verbalized Sampling across a range of tasks confirm the approach works — and the gains are not marginal. In creative writing tasks (poems, stories, jokes), VS increases output diversity by a factor of 1.6 to 2.1 times compared to direct prompting. Similar improvements appear in dialogue simulation, open-ended question answering, and synthetic data generation. Critically, these diversity gains do not come at the cost of factual accuracy or safety — both remain intact.
An additional pattern emerges in the data: more capable models appear to benefit more from Verbalized Sampling than less capable ones. This suggests that stronger base models have more suppressed diversity to unlock — their alignment training has constrained a richer underlying distribution, and VS provides a more effective key to open it.
Why This Changes the Conversation Around LLM Diversity
What makes this work analytically important is the reframing it offers. Most prior research treated mode collapse as an algorithmic problem — something to fix by improving RLHF methods, modifying reward models, or adjusting training objectives. This research repositions the diagnosis: the bottleneck is in the data, not the algorithm. Preference datasets, built by humans applying human cognitive patterns, carry structural biases that no amount of algorithmic refinement will fully remove if the underlying labels remain unchanged.
This data-centric perspective has broader implications. It suggests that the quality of human feedback — not just its quantity — is a fundamental constraint on how diverse and generative aligned models can be. For researchers and practitioners building preference datasets, the typicality bias finding is a concrete warning: annotator tendencies shape model personality in ways that are systematic, measurable, and consequential.
Verbalized Sampling, meanwhile, represents a practical inference-time answer to a training-time problem. Its value lies not just in the diversity improvements it delivers today, but in what it demonstrates: that the generative breadth of pre-trained models is not lost through alignment — it is merely suppressed, and addressable without starting from scratch.
FAQ
What causes mode collapse in large language models?
Mode collapse is primarily caused by typicality bias in preference data used during post-training alignment, where annotators systematically favor familiar-sounding text. This bias, rooted in cognitive psychology, trains models to converge on predictable outputs and avoid unusual but valid responses.
How does Verbalized Sampling mitigate mode collapse?
Verbalized Sampling prompts the model to generate a set of candidate responses and verbalize a probability distribution over them — for example, producing five versions of an answer alongside likelihood estimates. This approach reactivates the model’s pre-trained generative diversity without requiring any additional training or architectural changes.
In which applications does Verbalized Sampling improve diversity?
Verbalized Sampling improves output diversity across creative writing (poems, stories, jokes), dialogue simulation, open-ended question answering, and synthetic data generation — increasing diversity by 1.6 to 2.1 times over direct prompting in creative writing tasks alone.
Does Verbalized Sampling affect factual accuracy or safety?
No. According to the research, Verbalized Sampling improves diversity without sacrificing factual accuracy or safety, making it a viable option for production use cases where both variety and reliability are required.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Wall Street and crypto have been circling each other for years, but Morgan Stanley just made the relationship official in a very practical way. ETRADE crypto trading is now live for eligible U.S. clients, who can buy, sell, and hold Bitcoin, Ethereum, and Solana directly inside their existing brokerage accounts — no separate exchange account, no extra login, no fragmented portfolio view.
Key takeaways
ETRADE from Morgan Stanley has launched direct spot crypto trading for Bitcoin, Ethereum, and Solana for eligible U.S. clients.
A flat 0.50% commission per trade applies, with no additional spreads or markups; minimum trade is $10, maximum is $500,000.
Crypto accounts are powered by zerohash, which handles execution, liquidity, and custody.
Trading runs 24/7 on web and mobile, with both market and limit orders supported.
Crypto holdings are not protected by FDIC or SIPC; tax reporting uses IRS Form 1099-DA.
Morgan Stanley Launches Direct Crypto Trading on ETRADE
The move brings digital assets into the same brokerage interface where millions of Americans already manage their stocks, ETFs, and mutual funds. That might sound incremental, but the integration model matters more than it first appears. For a traditional investor who has spent years avoiding crypto because it meant opening an account on a separate exchange and navigating unfamiliar custody arrangements, this removes the friction almost entirely.
ETRADE now supports direct spot trading of Bitcoin, Ethereum, and Solana through a dedicated crypto account that links automatically to a client’s existing brokerage account. Funds move between the two accounts to support trades without requiring manual transfers in most cases. The result is a unified portfolio experience — one login, one view, three major cryptocurrencies sitting alongside equities.
24/7 trading with market and limit orders
Trading is available around the clock, seven days a week, on both the ETRADE website and its mobile app. Clients can place market and limit orders, and amounts can be specified in U.S. dollars or in coin quantity — including fractions down to eight decimal places. Support for Power E*TRADE platforms is expected to follow.
The trade size range is notably broad: a minimum of $10 makes crypto accessible to casual retail investors, while the $500,000 maximum per transaction gives more serious participants meaningful room to operate.
Service Features and Pricing Structure
Pricing is where ETRADE makes a deliberate statement. The platform charges a flat 0.50% commission on the notional trade value, with no additional spreads or markups layered on top. That transparent structure puts pressure on standalone crypto exchanges that have historically obscured costs inside wide bid-ask spreads.
zerohash powers execution, liquidity, and custody
Behind the scenes, zerohash handles the critical infrastructure: execution, liquidity sourcing, and secure custody. Clients link a dedicated zerohash-powered crypto account to their ETRADE brokerage account, or can open both simultaneously. zerohash maintains encryption and vulnerability management programs as part of its security posture, though the regulatory protections differ significantly from what investors receive on the traditional brokerage side.
ETRADE manages the client-facing experience, which means the interface, customer support, and portfolio tools remain consistent with what existing clients already know. That division of responsibility — brokerage UX from ETRADE, crypto infrastructure from zerohash — is a model that allows a major financial institution to move into digital assets without rebuilding its entire technical stack from scratch.
Regulatory and Security Considerations
Crypto holdings on ETRADE are not covered by FDIC or SIPC protections — a distinction that matters and that prospective users need to understand clearly. These are the insurance frameworks that protect bank deposits and brokerage accounts respectively, and they do not extend to digital assets regardless of where those assets are held. zerohash follows high security standards, but the absence of traditional deposit insurance is a structural difference from the rest of an ETRADE portfolio.
IRS Form 1099-DA for tax reporting
On the tax side, ETRADE will provide a 1099-DA form — the IRS’s dedicated reporting document for digital asset transactions — to report clients’ crypto activity. That standardized reporting is a meaningful operational convenience compared to manually tracking cost basis across multiple wallets and exchanges, which remains a common pain point for active crypto traders elsewhere.
Availability limited to eligible U.S. clients
The service is open to U.S.-based clients who meet standard ETRADE account requirements, but state-specific regulations mean eligibility may vary depending on where a client is located. International clients are not eligible at this stage.
What This Means for Retail Investors and the Broader Market
The strategic logic here is harder to miss. Morgan Stanley has offered crypto-related products to its wealth management clients for years and has previously introduced a spot Bitcoin ETF. The ETRADE launch extends that digital asset strategy directly into the retail mass market — a much larger and less sophisticated audience than the firm’s private wealth clientele.
For retail investors, the appeal is mainly convenience and familiarity. Portfolios can be viewed holistically through tools like Total Wealth View, and educational resources from Morgan Stanley covering market insights, long-term Bitcoin scenarios, and risk management accompany the rollout. The firm is clearly betting that lowering the technical and psychological barriers to crypto entry will translate into meaningful adoption among clients who have been curious but hesitant.
From a competitive standpoint, the flat 0.50% commission and the absence of hidden spreads creates a direct comparison point with established crypto exchanges. Whether that pricing holds up under competitive pressure — particularly if the token roster expands beyond the current three assets — will be worth watching.
Future enhancements may include crypto transfers into accounts and deeper wallet functionality, though those capabilities are not yet live. What is live already represents a significant shift: one of the largest financial institutions in the United States has placed Bitcoin, Ethereum, and Solana inside the same interface where everyday investors manage their retirement savings and equity portfolios. The infrastructure for mainstream crypto adoption just got considerably less complicated.
FAQ
Which cryptocurrencies can be traded directly on ETRADE through Morgan Stanley?
Eligible U.S. clients can trade Bitcoin, Ethereum, and Solana directly within their ETRADE brokerage accounts.
What are the trading hours and order types supported for crypto on ETRADE?
ETRADE supports 24/7 trading for crypto with both market and limit orders available on web and mobile platforms.
Are crypto holdings on ETRADE protected by FDIC or SIPC insurance?
No. Crypto holdings are not protected by FDIC or SIPC, but zerohash follows high security standards including encryption and vulnerability management programs.
What tax reporting documentation does ETRADE provide for crypto transactions?
ETRADE provides tax information via IRS Form 1099-DA for cryptocurrency trades.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.

Apples Klage wegen ihrer angeblichen Geschäftsgeheimnisse gegen OpenAI ist am 10. Juli wie eine juristische Granate eingeschlagen, und die Folgen sind erst am Anfang. Die Klageschrift – eingereicht vor einem Bundesgericht in Nordkalifornien – wirft nicht nur ein paar einzelnen fehlbaren Mitarbeitern vor, Dokumente in ihre Taschen geschmuggelt zu haben. Sie zeichnet vielmehr ein Bild von einer koordinierten, durchgängigen Operation, um vertrauliche Apple-Informationen zu erlangen, und nennt OpenAIs Chief Hardware Officer, Tang Tan, als Beklagten. Für ein Unternehmen, das eines der am meisten erwarteten Börsengänge in der Geschichte der Technologie ins Auge fasst, ist der Zeitpunkt kaum schlimmer zu treffen.

Es brauchte nur 207 Zeilen Code, um etwas zu tun, das die breitere Blockchain-Community für weitaus umfangreichere Arbeiten gehalten hatte. KaleidoSwap hat gezeigt, dass native RGB-Assets direkt auf der Liquid-Sidechain von Blockstream laufen können — ohne Bridge, ohne Custodian, ohne architektonische Überarbeitung. Die entscheidende Erkenntnis lag scheinbar im Verborgenen: Sowohl Bitcoin als auch Liquid verwenden ein identisches Taproot-Ausgabeformat gemäß BIP-341, was bedeutete, dass der Commitments-Mechanismus des RGB-Protokolls sich praktisch ohne Reibung in Liquid einfügen ließ.

Eine Bundesgroßjury hat eine Anklage wegen Krypto-Betrugs gegen einen Mann aus South Dakota erhoben. Ihm wird vorgeworfen, ein weitreichendes Anlagebetrugssystem mit einem Volumen von 20 Millionen US-Dollar betrieben zu haben, das Dutzenden von Opfern im gesamten Mittleren Westen leere Taschen und gebrochene Versprechen hinterließ. Der Fall wurde vom US-Justizministerium bekanntgegeben und rückt scharf in den Fokus, wie die Komplexität von Kryptowährungen gegen gewöhnliche Anleger eingesetzt werden kann.
Wichtige Erkenntnisse
Benjamin Paul Wiener sieht sich einer 29-fach erhobenen bundesrechtlichen Anklage gegenüber, darunter Wire Fraud, Geldwäsche, Bankbetrug und besonders schwere Identitätsdiebstähle.

Binance hat die Verzögerung der Aerodrome-Notierung von AERO um fünf Stunden verschoben und damit den Handelsstart von der ursprünglich angekündigten 11:00 UTC auf 16:00 UTC am 17. Juli 2026 verlegt. Für die Änderung wurde kein Grund genannt — lediglich eine kurze, direkte Mitteilung an Nutzer weltweit.
Wichtige Erkenntnisse
Der Handel mit Aerodrome (AERO) auf Binance sollte ursprünglich um 11:00 UTC am 17. Juli 2026 beginnen.
Der neue Handelsstart wurde auf 16:00 UTC am selben Datum verschoben — eine fünfstündige Verzögerung.
Binance veröffentlichte das Update als allgemeine Bekanntmachung, die an alle Nutzer weltweit gerichtet war.

Der Merck-Kurs stieg am 16. Juli stark, nachdem die FDA Lipfendra zugelassen hatte, ihren oralen PCSK9-Inhibitor gegen Hypercholesterinämie. Die Aktien von MRK sprangen von 124,33 $ auf 127,63 $ und schlossen bei 127,63 $. Die Struktur des Tagescharts hatte sich bereits verbessert. Der Auslöser schärfte lediglich die Kante.
MRK — Tageschart mit Kerzen, EMA20/EMA50 und Volumen.
Wichtige Erkenntnisse
Der Merck-Kurs sprang von 124,33 $ auf 127,63 $ am 16. Juli, nachdem die FDA Lipfendra zugelassen hatte, einen oralen PCSK9-Inhibitor zur Behandlung von Hypercholesterinämie.
Tägliches EMA-Stacking — EMA20 bei 123,78 $, EMA50 bei 120,78 $, EMA200 bei 110,05 — bestätigt einen klar bullischen Trend für MRK.