WHALE WATCHER

Gergely Orosz showed Jev pricing on his new search site with a $0 bill after it handled more than 1 million tokens.
Key Takeaways
Gergely Orosz showed a $0 bill for more than 1 million tokens on his new search site
Orosz posted the figure on X and said Jev pricing is “per billions of tokens not millions,” while OpenAI’s listed pricing shows separate charges for input and output tokens.
Search is token-hungry, with prompts, retrieval steps, ranking logic and final answers all adding up
Jev must show whether the same economics hold across larger workloads, steadier traffic and harder tasks
Orosz posted the figure on X and described Jev’s billing as “per billions.” OpenAI‘s listed pricing shows separate charges for input and output tokens.
Jev Pricing Turns A Tiny Bill Into A Bigger AI Search Pricing Fight
The $0 screenshot is small in dollar value but hits a pressure point for AI Search products because search is token-hungry. Prompts, retrieval steps, ranking logic and final answers all add up.
Many teams have learned model quality is only part of the buying decision.
The harder question is whether a tool stays cheap when real users query it all day. A service that keeps those costs low can offer free tiers, absorb experimentation and delay price hikes, giving builders room to iterate.
Orosz’s example highlights how little most users see of model economics.
A token is a small chunk of text, not a whole word, and vendors charge for the total number processed. Passing 1 million tokens at no visible cost suggests either aggressive subsidies, very low per-token rates, heavy caching, or a billing floor that rounds small usage down.
Also Read: AI Hardware Startup Positron Wins A Buyer That Promises To Come Back For More
From Premium Model Bills To Jev Pricing Pressure On AI Search
Products with many short queries can face costs that rise with usage.
A search tool, coding assistant or support bot can look cheap in testing and turn expensive when usage climbs, so developers hunt for predictability as much as raw performance.
If a provider bills by the billion tokens rather than smaller blocks, it changes how teams think about margin. Small customers may not care about shaving fractions of a cent per query, but at product scale those fractions pile up fast.
Pricing has become a product signal alongside benchmarks and longer context windows.
The newer fight is whether a model can be deployed widely without turning each feature into a budget line item. For independent builders, a $5 surprise bill is far more meaningful than it is for Microsoft or Google, making AI Search cost mechanics especially relevant to solo developers and small teams.
What Jev Pricing Still Needs To Prove
One screenshot does not settle the question.
Jev must show whether the same economics hold across larger workloads, steadier traffic and harder tasks. Cheap pricing can come from efficient inference, architectural choices like retrieval and routing, or from promotional tactics that do not last.
Buyers will want to know whether Jev is cheap because it is efficient, because it routes work differently, or because it is eating margin to win share.
A lower token bill only helps if answers stay useful, since quality tradeoffs matter when search systems mix models with retrieval.
How Per-Billion Billing Changes The Meter For AI Search
If Jev bills in very large token blocks, while OpenAI and many rivals quote prices in smaller units, then a developer who crossed 1 million tokens and still saw $0 suggests Jev’s meter may not tick up for light usage the way standard API billing does.
Orosz said Jev bills per billions of tokens, not millions, and said an LLM would have charged $3 to $5 already for more than 1 million tokens on his new search site. That helps explain why the bill still showed $0.00.
For search apps that fire many small calls, treating those calls as tiny fractions of a billion-token block lets teams test features without watching pennies on every query. The next things to watch are fuller Jev rate cards or case studies.
Until then Saturday’s post is a marker that pricing, not just model quality, is the next battleground in AI Search.
In his post, Orosz said an LLM would have charged $3 to $5 already for that usage, while Jev bills per billions of tokens rather than per millions. That pricing detail helps explain why more than 1 million tokens could still show a $0.00 bill.
Read Next: Supermicro Hands Companies A Memory Saver For Thousands Of Chats

IBM opened its most advanced quantum system in Chicago and linked it to a three-way error-correction effort with AMD and Oak Ridge National Laboratory.
Key Takeaways
IBM opened its most advanced quantum system in Chicago
IBM paired the Chicago site with a new error-correction co-design effort involving AMD and Oak Ridge National Laboratory
Oak Ridge National Laboratory lists “Future of Computing,” which includes quantum, on its homepage.
IBM offers quantum computing products and services.
Oak Ridge National Laboratory lists “Future of Computing,” which includes quantum, on its homepage.
IBM offers quantum computing products and services.
Chicago Puts IBM Quantum Computer On Display
The immediate news is not that quantum computing solved its biggest problem. It did not.
The news is that IBM is spending political, engineering and partner capital around one location and one bottleneck, rather than treating the machine as a lab curiosity.
That matters because most quantum systems still fail after short runs. Their qubits are fragile, noisy and easy to disturb.
A more powerful machine without better error handling can produce more wrong answers faster.
IBM’s release said the Chicago computer marks a milestone for Illinois’s quantum future. It also linked the Chicago opening to a co-design platform with Oak Ridge and AMD.
That combination matters more than the ceremony. Quantum hardware builders increasingly need classical chip companies and national labs to handle the heavy math and control loops that stable quantum computing demands.
From Raw Qubits To Useful Machines
The backdrop is a long industry shift away from headline qubit counts alone.
Earlier quantum announcements often centered on how many qubits a machine had. Over the last two years, the focus has moved toward fidelity, calibration, connectivity and error suppression.
That shift mirrors what happened in AI infrastructure.
Bigger models alone stopped being enough once users wanted reliable outputs, lower latency and lower operating cost. In quantum computing, the comparable hurdle is not training data or inference speed.
It is whether the machine can preserve information long enough to finish a calculation.
IBM has been pushing that message for several product cycles. Friday’s Chicago opening extends that strategy by anchoring the story in a physical site and in partnerships that attack the control stack around the qubits, not just the qubits themselves.
Also Read: GitHub Guide Puts Copilot Across 4 Coding Jobs, Humans Decide
IBM Quantum Computer Faces The Error Wall
The central problem is error correction.
Quantum bits can flip or lose their state from heat, vibration, radiation or imperfections in the hardware. Because reading a qubit can disturb it, engineers cannot simply check it the way they would check a memory chip in a normal server.
Instead, companies like IBM spread information across groups of physical qubits and use repeated measurements on related states to infer whether an error happened.
That requires a large amount of classical computation around the quantum chip. It also demands very fast decisions, because the correction logic must react before the quantum state decays.
AMD enters that picture as a classical computing partner.
Oak Ridge brings national lab expertise in simulation and high-performance computing. IBM brings the quantum hardware and software stack.
If that loop gets faster and more accurate, developers get closer to systems that can run longer jobs with fewer useless outputs.
What The Chicago Site Changes For Partners
For enterprise buyers and research groups, the Chicago site is a sign that IBM wants quantum adoption to look more like a durable infrastructure program and less like a series of disconnected demos. A local cluster of hardware, lab relationships and chip partners can shorten test cycles.
The next thing to watch is whether IBM publishes performance data that shows the Chicago system or the co-design platform cuts real error rates or extends circuit depth.
Without that, the opening remains an infrastructure milestone, not a proof of quantum advantage.
How AMD Fits IBM’s Error Fight
IBM’s Friday release named AMD and Oak Ridge National Laboratory in a co-design platform for quantum error correction, and that pairing says a lot about where the bottleneck sits. Quantum processors do not work alone.
Each qubit needs control electronics, calibration routines and fast classical feedback that decide whether the machine should keep going or correct course.
For AMD, this matters because quantum control is a new demand source for advanced classical compute. The job is not training a model like OpenAI or Meta would.
It is running tight loops around a quantum chip, often with strict timing limits, so the system can detect faults before the quantum information fades.
For labs and enterprise developers using IBM Quantum, the practical question is simple. Can the surrounding stack catch more errors without slowing the whole machine so much that the gain disappears.
That is the tradeoff to watch as IBM, AMD and Oak Ridge move from announcements to published results.
Read Next: GitHub Guide Puts Copilot Across 4 Coding Jobs, Humans Decide

Security company Apiiro found 2,539 real secrets across 1,994 public GitHub repositories, including credentials in Claude Code settings files, git history, shared settings and agent transcripts.
Key Takeaways
Apiiro found 2,539 real secrets across 1,994 public GitHub repositories, including credentials in Claude Code files and related agent records.
The issue centers on “.claude/settings.local.json” local files created after using the “don’t ask again” option in Claude Code
Apiiro said it identified real secrets in public repos, meaning cloud keys, API tokens or internal credentials may have been exposed
Teams should review ignore rules, extend secret scanners into hidden folders, and rotate any key that reached a public repository
Apiiro said the problem centers on “.claude/settings.local.json,” a local file developers can accidentally commit after using a “don’t ask again” option inside Claude Code. Out of 104,185 public copies of that file on GitHub, 1,122 held a real secret, and 63 organizations were affected, including one that confirmed its key had been live.
Also Read: Anthropic's Claude Code Has Been Sending Users' Email Addresses To Websites
The numbers matter because these were not theoretical test strings. Apiiro said it identified real secrets in public repos, which means cloud keys, API tokens or internal credentials may already have been exposed to anyone scraping GitHub.
That turns a convenience feature into a supply chain risk.
Claude Code Secrets Turn Convenience Into Exposure
Coding agents increasingly ask for standing permission so they can run commands, install packages or call tools without stopping for every approval. That makes the tools feel faster, but it also creates a new class of residue where sensitive settings end up written to disk and later pushed into version control.
Once a file enters a public repository, copies can spread quickly through forks, mirrors and search indexes.
Even if a developer deletes the file later, the secret may already have been copied.
Apiiro’s finding lands in a wider shift from autocomplete toward agentic coding. The newer tools do not just suggest lines of code, they also inspect files, execute tasks and remember decisions.
That broader scope makes local configuration files far more sensitive than they were in older coding assistants.
From Helpful Defaults To Real Attack Surface
Security teams have warned for years about hardcoded keys in source code. The new twist is that AI agents can create those exposures through permission memory rather than a developer manually pasting credentials into an app.
That distinction matters for response plans.
Traditional secret scanning looks for tokens inside source files, while agent-generated risk can sit in hidden local paths, tool settings or machine-specific folders that developers never expected to publish. Security programs built around code review alone can miss that path.
The problem also fits a pattern in enterprise AI rollouts.
Companies add agents to speed engineering work, then discover the tool created new storage, logging or approval states outside normal policy controls. Those blind spots are one reason many organizations now treat AI developer tools as security products, not just productivity software.
What Developers Need To Check Now
Teams using Claude Code should review whether local settings files are excluded from commits and whether repository scanners cover hidden directories.
They should rotate any key that may have reached a public repository, because deletion does not guarantee the credential stayed private.
Apiiro’s report specifically highlights “.claude/settings.local.json” as the path where persistent permission records are stored after using the “don’t ask again” option. That file is the vector tying Anthropic‘s agent behavior to exposed tokens in public GitHub repositories.
Developers should watch for vendor changes to file paths, ignore rules and permission storage because those interface details decide whether agent speed stays local or becomes a public leak.
Why Hidden Files Matter More
Claude Code secrets are dangerous because they hide in places many teams do not review closely.
A developer may think the sensitive action happened inside the assistant, but the real evidence lives in a config file on disk. When that file gets committed to GitHub, the exposure stops being a personal mistake and becomes a repository problem.
For platform teams, the fix is straightforward.
Add the local file path to ignore lists, extend secret scanners into hidden folders and block commits that contain known agent settings files. Rotate credentials fast when a repo goes public.
For vendors like Anthropic, pressure will grow on default storage choices and safer permission memory, the next battleground in coding agents may be where the agent writes its memory.
Read Next: GitHub Guide Puts Copilot Across 4 Coding Jobs, Humans Decide

Broker Axi said on Friday it opened pre-IPO products tied to OpenAI and Anthropic, letting retail traders bet on price moves without owning real shares or shareholder rights.
Key Takeaways
Axi opened pre-IPO trading products tied to OpenAI and Anthropic on its platform
The contracts give traders synthetic exposure without owning actual equity or shareholder rights
Axi did not disclose contract sizing, pricing mechanics or liquidity sources
Brokers must map private valuations, secondary indications or internal pricing models into tradeable instruments, which can widen spreads
Axi said on Friday that it had opened the products on its platform and described them as pre-IPO trading opportunities tied to OpenAI and Anthropic. The broker did not disclose contract sizing, pricing mechanics or liquidity sources, according to the announcement.
Anthropic already filed for an IPO on Oct. 26, 2026, seeking a Nasdaq listing under the ticker ANT, Axi said. OpenAI is also studying a possible IPO over the longer term.
Pre-IPO OpenAI access on a retail broker usually means synthetic exposure, often through contracts for difference or internal reference pricing, not a legal claim on company stock. Traders can win or lose on price movements while staying outside the cap table.
Pre-IPO OpenAI And Anthropic Is A Synthetic Bet
Axi is an online foreign exchange and CFD broker.
In that structure, the customer and broker settle the difference between an opening and closing price instead of transferring the underlying asset.
For private companies, that creates a wrapper with no continuous public market for OpenAI or Anthropic shares. A broker must therefore map private valuations, secondary indications or internal pricing models into a tradeable instrument.
That can widen spreads and make price discovery harder than in listed stocks.
From Private Rounds To Retail Demand
The new contracts build on a broader retail scramble for private AI exposure. OpenAI and Anthropic have drawn strong market attention before reaching public markets.
That gap creates demand for stand-ins.
Some investors chase venture funds. Others buy suppliers or cloud names.
Axi’s product takes the most direct route by letting traders speculate on a reference value for the private company itself, even though no common shares change hands.
Also Read: Anthropic Guest Post Says AI Can Speed Scientific Model Work
The Risk Sits In The Wrapper
The biggest risk is not only whether OpenAI or Anthropic becomes more valuable. It is whether the contract price fairly tracks any real secondary market.
If the broker is the main pricing venue for a retail user, slippage and spread costs can matter more than the headline thesis.
There is also event risk. A funding round, tender offer or governance change can reset private valuations in chunks instead of the minute-by-minute pattern equity traders expect.
That makes mark-to-market swings on a synthetic contract harder to interpret.
What Traders Should Watch Next
If pre-IPO OpenAI and Anthropic contracts draw volume, other brokers will likely copy them. Key facts to watch are margin terms, overnight financing charges and how often the reference price updates.
Those details decide whether the product behaves like a genuine valuation proxy or an expensive short-term punt.
For the AI market, the product shows how far private lab branding has traveled into retail finance. OpenAI and Anthropic have become tradeable narratives before they have become public companies.
How Axi’s Contract Works Without Real Shares
With Axi’s pre-IPO OpenAI product, the buyer is not purchasing stock in OpenAI.
The trader is entering a derivative with Axi. If the broker’s reference price rises, the trader can profit.
If it falls, the trader takes the loss.
That matters for retail users because shareholder rights do not come with the contract. There is no vote, no transfer record and no guaranteed link to a future IPO allocation.
The position is only as good as the broker’s pricing method and its ability to honor gains and losses.
The same issue applies to Anthropic. Private-company prices move in large steps around funding rounds, employee sales or tender offers.
A synthetic instrument tries to smooth that into a live market. For active traders, that creates access.
For long-term investors, it creates tracking risk. Before these products spread, traders should watch whether brokers publish their valuation inputs and how those inputs change after the next financing event.
Read Next: OpenAI Launches BioBasicsBench As GPT-6 Astra Hits 66.7%

Four small Nvidia DGX Spark desktop boxes just served the open DeepSeek V4.1 Flash model to 16 different users at once, at about 24 tokens a second each, with no server rack, no liquid cooling and no cloud bill.
Key Takeaways
A custom, open-source build of DeepSeek V4.1 Flash served 16 different users at 380 tokens a second in total on four linked DGX Spark boxes, and 438 tokens a second when 16 copies of one request ran together
A single user got 80 tokens a second in chat and 138 tokens a second when writing code
The setup handled a 1 million token context and read prompts at 10,000 tokens a second
It beat SGLang TP4 on the same boxes by 18% to 25% for one user, with about 1.9x faster prompt reading
What The Test Showed
A developer posting as Picomoorhead open-sourced the build, a serving engine written for this one model on four DGX Spark units. Tokens are the word chunks AI models read and write, so 80 tokens a second is faster than anyone can read.
Industry analyst Patrick Moorhead, founder of Moor Insights &amp. Strategy, shared the numbers and said they answer the main complaint about DGX Spark. “The knock on DGX Spark was that it’s a developer box, not an inference box,” he wrote. “This open-source DeepSeek V4.1 Flash custom build on four of them says otherwise.”
Four DGX Spark boxes running DeepSeek V4.1 FlashResult Chat, one user80 tokens/sec Code, one user138 tokens/sec 16 different users at once380 tokens/sec (~24 each)16 copies of one request438 tokens/sec (~27 each) Prompt reading speed10,000 tokens/sec Context window1 million tokens Vs. SGLang on the same boxes18-25% faster, ~1.9x prompt reading
Also Read: Amazon Adds Nvidia Chip Sharing To Cut AI Serving Costs
Software Was The Bottleneck
Moorhead said the result came from software, not new hardware. “The hardware was never the limiter. The software was,” he wrote, adding that an engine built for one model on one setup beats general-purpose serving tools.
The whole rig is four desktop boxes and a 200G network switch. Moorhead called it “the template for departmental AI,” meaning open models from DeepSeek, small Nvidia boxes and inference software written by the open-source community.
Why It Matters For Companies
Nvidia pitched DGX Spark as a desk-side machine for developers to build and test AI models, while serving answers to many users was left to data center GPUs or the cloud. Serving 16 people at once on four desktop units points to a middle option for teams that want to keep sensitive prompts in house or cap cloud spending.
The catch is that this engine is custom-built for one model on one setup. Teams that switch models or hardware would need similar tuning to get the same speed.
What To Watch Next
The X post said image support was coming soon, and the GitHub README says images work through the normal OpenAI API. The next test is whether other teams can repeat the numbers with the open-source code, and whether similar engines appear for other open models.
Read Next: Small Investor Blockmate Bets On Nvidia Server To Cash In On AI

Google wird wegen der Vorwürfe verklagt, Gemini habe einem Teenager auf einem Schul-Chromebook nach Tausenden von Nachrichten mit dem Chatbot dabei geholfen, einen Suizidversuch vorzubereiten.
Die wichtigsten Erkenntnisse
In einer diese Woche eingereichten Klage wird behauptet, Googles Gemini habe einem Teenager auf einem Schul-Chromebook nach Tausenden von Chatbot-Nachrichten dabei geholfen, einen Suizid vorzubereiten
The Hill berichtete, Google habe später erklärt, dass Gemini, Gemini Notebook und die Suche für Schulen standardmäßig deaktiviert sein werden und Schulen auswählen können, ob sie einen, zwei oder alle Dienste aktivieren möchten.

GitHub said Copilot can help with four kinds of work including planning, writing code, reviewing changes and investigating production issues, while developers still make the final call.
Key Takeaways
GitHub used “agentic engineering” to describe workflows where Copilot helps with planning, drafting, review prep, and production investigation.
The guide lists four kinds of work where Copilot can help. Planning, writing code, reviewing changes and investigating production issues
A second page shows Copilot now spans chat, pull requests and task flows inside the development system
The post says developers still review tradeoffs and decide whether a change ships.
The post said the point is not just faster typing. The company runs the code hosting platform behind Copilot, and the post frames AI as a helper across the software loop rather than a bot that finishes a feature on its own.
GitHub Agentic Engineering Expands Copilot’s Job
The guide lists four kinds of work where Copilot can help: planning, writing code, reviewing changes, and investigating production issues.
The post says developers still review tradeoffs and decide whether a change ships.
A second page shows how Copilot is built into different parts of GitHub.
Also Read: GitHub Gives All Copilot Users OpenAI Tool That Solves More Code
From Autocomplete To Workflow Coverage
Earlier coding assistants sold speed at the point of writing and worked for boilerplate, but they left large parts of the engineering day untouched, especially triage, review and debugging after release.
The guide’s framing follows a wider industry turn toward agents that can inspect codebases, read issue threads and propose multi-file changes, and GitHub is trying to keep Copilot central as those tools spread beyond editors into repositories, CI pipelines and production support.
The Human Gate Stays In Place
The guide says developers still decide what ships. The guide says Copilot can help investigate production issues, while developers still review changes and decide what ships.
In software terms, agentic engineering means distributing small pieces of reasoning across the workflow while ensuring humans approve changes that go to production.
How Agentic Engineering Changes Daily Work
The guide uses agentic engineering to describe software work where Copilot participates in several stages instead of one.
In a normal team, a developer may start with an issue, ask Copilot for an implementation plan, generate a draft, request a review summary and then query logs after a failed deploy. It packages those separate moments into one work pattern, reducing switching between tools and time spent reconstructing context.
For engineering managers, the change is about process design.
A team using Copilot can assign AI to repetitive steps while keeping people on approval and system design.
Read Next: Supermicro Hands Companies A Memory Saver For Thousands Of Chats

Das US-Finanzministerium verhängte Sanktionen gegen acht kanadische Wechselstuben, die laut seinen Angaben iranische Rial in Kryptowährungen umtauschten und Gelder mit Verbindungen zu Irans Revolutionsgarden transferierten. Mehrere von ihnen wuschen zudem Drogengeld für den ehemaligen olympischen Snowboarder und späteren Drogenhändler Ryan Wedding.
Die wichtigsten Erkenntnisse
Die Sanktionsbehörde des US-Finanzministeriums nahm am 6. November im Rahmen der Operation Economic Outcast acht in Kanada ansässige Geldtransferunternehmen ins Visier
Das Finanzministerium erklärt, die Unternehmen tauschten iranische Rial gegen digitale Vermögenswerte und arbeiteten mit den sanktionierten iranischen Kryptobörsen Nobitex, Wallex, Ramzinex, Bitpin und Shelbit zusammen

Polymarket, a website where people bet money on real world events, is investigating a trader who put nearly $200,000 on a Nebraska Senate race an hour before a poll and later made more than $77,000.
Key Takeaways
A trader reportedly placed nearly $200,000 on a Nebraska Senate market before a poll on Oct. 31 and later netted more than $77,000
Cointelegraph reported the bet was placed about an hour before a Nebraska Senate poll became public and Polymarket said it was investigating
Vanderbilt University researchers said they saw a pattern in the trades that was hard to dismiss as random
Cointelegraph said the same user also placed large bets on Kalshi, the federally regulated U.S. prediction exchange
Cointelegraph reported that Polymarket was investigating unusual activity tied to the Nebraska Senate race.
A Senate Poll Bet Draws Scrutiny
The reported trade centered on a market about Nebraska Senator Pete Ricketts‘ reelection. Cointelegraph said Vanderbilt University researchers saw a pattern that was hard to dismiss as random, even if there is no public proof the trader held nonpublic polling data.
That distinction matters because prediction markets invite aggressive bets by design, but the line between informed analysis and illegal use of confidential information is where regulated finance usually steps in, and allegations of Polymarket Insider Trading hinge on that difference.
Polymarket Insider Trading Tests Crypto Markets
Prediction markets sell themselves as cleaner signals than social feeds or pundit panels because prices move with money, not just opinions.
That promise weakens if traders believe some wallets get first look at private polls, government data or campaign research. A market can still settle correctly and still be unfair on the way there, which is why surveillance matters as much as volume.
Also Read: Supermicro’s Legal Bills Jump 549% As SEC And Export Probes Drag On
The Polymarket Insider Trading questions deepen when researchers say timing patterns are unlikely to be random. Cointelegraph said the same user also placed large bets on Kalshi, the federally regulated U.S. prediction exchange, which means scrutiny may not stay confined to one crypto venue.
From Venezuela Case To Election Bets
This is not the first time prediction markets have faced insider-trading questions.
That earlier case was more direct because prosecutors tied the trades to operational knowledge inside a government action, while the new episode is murkier but revives the same basic question about whether fast-moving event markets can police information asymmetry before confidence breaks, a core concern when considering Polymarket Insider Trading allegations.
How Prediction Markets Spot Suspicious Timing
A trader who buys before new information becomes public can lock in a profit once everyone else reprices the odds, which makes timing the core surveillance problem, not just trade size.
For researchers and compliance teams, suspicious activity often starts with a clock, if a wallet buys heavily minutes before a poll drop, a court filing or a military action, Polymarket and Kalshi can compare timestamps, counterparties and any repeated pattern across markets. The public shows the price path, but the platform holds the deeper order data that can reveal whether one account keeps arriving just before the market moves.
For cryptocurrency readers, this matters because event contracts look more like financial instruments each year. If Polymarket wants institutions, campaigns or media-linked markets to grow, it will need controls that resemble exchange surveillance, not just wallet transparency. The next signal to watch is whether the company names the market, freezes profits or refers findings to authorities, and whether those steps address concerns of Polymarket Insider Trading.
Read Next: AWS Uses Okta Logins To Stop AI Agents Going Rogue At Work
The New York Times also reported that Vanderbilt researchers identified suspicious prediction market activity before the release of two recent Times/Siena polls, not just the Nebraska Senate market. The same user also bet $622,608 on Michigan’s Senate race before another poll release and made more than $180,000.

Amazon Web Services laid out a way for companies to tie every AI agent action back to a real employee, using Okta sign-ins so a work bot can only touch the tools that person approved, one call at a time.
Key Takeaways
AWS showed how Okta Cross App Access can link an AI agent’s rights to a named user instead of a broad service account
Bedrock AgentCore Gateway redeems the Okta grant before the agent can reach a protected tool
Cedar policies check every single tool call, and AgentCore Observability logs the trail
A follow-up guide extends the setup to multiple agents, a central gateway and agents running on Amazon EKS
AWS explained the setup for a single agent first, then expanded it to multiple agents, a central gateway and observability hooks.
AI Agent Identity Moves Past Demo Mode
In the basic version, one agent acts on behalf of a named user. Okta Cross App Access issues a grant, and Amazon Bedrock AgentCore Identity redeems that grant before the agent touches any downstream tool.
The bigger version adds a central gateway, a second agent on Amazon EKS and a policy check for each tool call. Amazon said the pattern gives enterprises a single front door for AI agent identity, instead of separate service accounts spread across apps.
That matters because AI agents are shifting from chatbots to software that acts. Once an agent can open tickets, search internal files or trigger workflows, the hard problem is no longer only model quality. It is deciding which agent may do what, for which user, and under what limit.
Also Read: OpenAI Sets 5 Rules To Stop AI Agents Going Off Script
Where Agent Permissions Break First
The risk shows up when teams bolt agents onto old app permissions. A user may have rights inside Salesforce or Workday, but a background agent often runs under a broad service identity, which blurs who actually approved an action.
The AWS design narrows that. Okta issues the user-linked authorization, Bedrock AgentCore Gateway receives it, Cedar, Amazon’s policy language, decides whether that exact tool call is allowed, and AgentCore Observability records the event trail.
That gives security teams, compliance officers and developers at regulated companies what they have been missing: agents that can act, plus proof that each action maps back to a user-approved chain.
Why This Beats The Service Account Shortcut
The key change is that Bedrock AgentCore does not treat the agent like a generic backend script. It treats the agent as a separate actor that inherits limited rights from the person it works for.
For companies already on Okta, that avoids the common shortcut of giving one service account broad standing access across many systems, a setup that makes it hard to trace who signed off when something goes wrong.
The next thing to watch is whether AWS turns this documented pattern into default settings inside Bedrock AgentCore. If it does, user-linked agent identity could move from architecture guidance to a standard enterprise control.
Read Next: EU Starts Grilling AI Labs After OpenAI Agents Escaped And Hacked A Company

Das KI-Unternehmen Anthropic erklärte, dass KI-Systeme Wissenschaftlern dabei helfen können, wissenschaftliche Modelle zu erstellen, auszuführen, anzupassen und zu überarbeiten, um zu prüfen, welche Erklärungen durch die Daten gestützt werden.
Die wichtigsten Erkenntnisse
Anthropic veröffentlichte auf X ein Forschungsupdate, in dem es hieß, dass KI Forschenden helfen kann, wissenschaftliche Modelle als Code zu schreiben, auszuführen, anzupassen und zu überarbeiten.
Der Beitrag hebt die Arbeit der Geförderten Jakob Macke, Daniel Gedon und Stefan Wahl hervor
Anthropic veranschaulicht anhand des Hodgkin-Huxley-Modells einer Nervenzelle, wie Wissenschaftler wissenschaftliche Modelle erstellen, einschränken und interpretieren.
Anthropic stellt KI als Werkzeug für die Entwicklung wissenschaftlicher Modelle dar und nicht nur als Schreibhilfe.

Amazon Web Services added Nvidia‘s chip-sharing system to SageMaker so several models on an endpoint can run on one graphics chip at once, which can lower the cost per request.
Key Takeaways
SageMaker Inference now supports Nvidia Multi-Instance GPU (MIG) partitions so different models on an endpoint can share one accelerator
AWS added two optional API fields to let customers place an endpoint on either a full GPU or a specific MIG slice
MIG partitions give each slice its own memory and compute allocation for isolated hardware-level resources
Packing mixed workloads onto one accelerator can raise utilization and lower the cost per request for inference fleets
AWS said in a Friday blog post that SageMaker Inference now supports Multi-Instance GPU for model endpoints. Nvidia describes MIG as hardware partitioning that turns one physical GPU into several isolated instances.
Amazon Web Services runs cloud infrastructure and machine learning products, and SageMaker Inference is its managed system for serving models through production endpoints.
The new support targets a basic economics problem in AI serving, small and medium models often do not use a full modern GPU.
One GPU, More Paying Work
A dedicated endpoint gives a model predictable performance, but it can also strand expensive compute. If a model only needs a fraction of a GPU, the unused portion still gets reserved and billed.
MIG changes that by slicing the chip at the hardware level. Each slice gets its own memory and compute allocation, which means a small model can share a GPU with another model without software-only sharing tricks.
That matters most for inference fleets with mixed workloads. A company may serve a chatbot, a classifier and an embedding model at once, all with different traffic patterns.
Packing them onto one accelerator can raise utilization and lower the cost per request.
From Training Race To Serving Margin
The AI market spent two years talking about buying more GPUs. The next fight is getting more revenue from the GPUs already in racks.
Inference, not training, is where many cloud bills settle into recurring operating cost.
Also Read: Amazon Says New Memory Trick Could Cut Chatbot Costs
Friday’s update follows a broader shift across the cloud stack. Instead of assuming one large model per accelerator, providers are tuning around fragmentation, isolation and scheduling.
Those are less flashy problems, but they decide whether AI products can hold margin at scale.
AWS did not tie the feature to one model family. That gives it wide relevance for teams deploying image models, smaller language models or task-specific systems that do not justify exclusive access to a top-end chip.
How SageMaker Uses Nvidia Multi-Instance GPU
SageMaker Inference manages the servers that answer model requests after training ends. A model endpoint is the live service that receives prompts or data and returns predictions. With MIG enabled, AWS can place an inference component on a defined fraction of a compatible Nvidia GPU instead of the whole device.
The key change is isolation at the hardware layer.
A software scheduler can time-share a GPU, but one noisy workload can still interfere with another. MIG splits resources before jobs run, so each slice has reserved capacity and memory.
That makes latency and failure domains easier to reason about.
For startups and internal platform teams, the gain is not just lower cost. MIG also makes it easier to match hardware to workload size.
A lightweight moderation model no longer needs to ride alone on the same class of accelerator used for far larger systems.
Operational Complexity And Placement Decisions
GPU partitioning can cut waste, but it also adds placement choices. Teams now need to decide when to use a full GPU, when to use a slice and how to rebalance as traffic changes.
That makes observability important.
If an endpoint outgrows its slice, latency will rise and the savings case disappears. The winning setups will be the ones that connect model size, traffic and partitioning policy without forcing engineers to micromanage every endpoint.
Read Next: Nvidia Says 8,000 Institutions Now Use Its 6G Research Tools

Tencent, one of China’s biggest tech firms, said its new AI system Hy4 earned the highest overall score in an evaluation based on 579 real-world finance tasks, ahead of DeepSeek-V4.1-Flash-1017.
Key Takeaways
Hy4 scored 3.04 across 579 real-world finance tasks, Tencent said
Tencent said Hy4 improved by 12.03 points on another measurement in a Thai release
Tencent named DeepSeek-V4.1-Flash-1017, Kimi K3.1 Pro and GLM-5.3 as comparison points
Tencent positioned Hy4 as open source and said the 579 tasks are drawn from real financial work
Tencent said the release puts Hy4 among the leading open-source AI models.
Tencent Hy4 Targets Finance Work
Tencent Hy4 matters because the benchmark is not framed as a general chatbot bakeoff, and Tencent said the 579 tasks are drawn from real financial work, which is where model buyers care about precision, document handling and multi-step reasoning more than style.
Tencent compared Hy4 with DeepSeek-V4.1-Flash-1017, Kimi K3.1 Pro and GLM-5.3 on a finance-focused evaluation.
Tencent is one of China’s largest technology groups and has been expanding its model lineup as open-weight AI turns from research branding into enterprise distribution.
Also Read: AWS Case Study Says Exowatt Slashes AI Bills And Beats Big Rivals
A China Open Model Fight Is Getting Tighter
Tencent Hy4 lands in a crowded field, and the release itself names DeepSeek-V4.1-Flash-1017, Kimi K3.1 Pro and GLM-5.3 as comparison points. That is the background readers need for Friday’s claim. Chinese model makers spent much of the last year chasing cheaper inference, stronger reasoning and local enterprise adoption.
Tencent said Hy4 joins the ranks of the world’s leading open-source AI models. The release says Tencent Hy plans to release the model weights and system card in the coming days, and it says Hy4 is available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom AI Assistant, ima, and APIs on OpenRouter, Hermes, and OpenCode.
Why 579 Tasks Matter More Than A Chat Demo
Tencent Hy4 is being sold on test coverage rather than a glossy consumer feature, and a 579-task finance benchmark gives buyers a broader sample of performance than a few staged prompts.
Tencent said the evaluation used 579 real-world finance tasks and that Hy4 was trained on scenarios tied to fund research, securities analysis, and quantitative investment strategy.
How Open Weights Change Buyer Control
Open-weight models give companies a different operating model from API-only systems. With a closed API, a bank sends prompts to a remote service and accepts the vendor’s pricing, latency and update schedule, whereas with an open model a firm can run the weights on its own hardware or cloud account, test changes internally and tune the system for a narrow workflow.
Tencent said Hy4 was the top-performing open-source model in its blind evaluation based on 579 real-world finance tasks.
Tencent’s release names the scores, but buyers will want deployment details, licensing terms and proof that Hy4 keeps its edge outside a curated finance benchmark.
Read Next: OpenAI Launches BioBasicsBench As GPT-6 Astra Hits 66.7%
Tencent today announced Hy4 launched with visual understanding and new computer-use capabilities, plus improvements in instruction following, long-context comprehension and long-horizon agentic tasks. The finance benchmark win was one result within that broader release.

Stablecoin giant Circle says more than 100 banks, payment firms and apps are now building on Arc, its own blockchain, and over $30 billion in USDC has already moved across it since launch.
Key Takeaways
Circle said 100+ banks, payments leaders and apps are building on Arc, including BNY, HSBC, Mastercard and Visa
More than $30 billion in USDC has moved on Arc since its mainnet went live on September 16
Circle shared the figures during its third-quarter earnings livestream on Friday
Fees on Arc are paid in USDC, so users never need to hold a separate volatile token
What Circle Said
Circle posted the update on X on Friday, during its third-quarter earnings livestream. “Arc is live, with 100+ banks, payments leaders, and apps building on the network,” the company said, naming BNY, HSBC, Mastercard and Visa.
The company added that more than $30 billion in USDC has moved on Arc since its mainnet launch. It did not say how much of that came from banks versus apps, or which of the named partners are live in production.
Also Read: Tether's $8B Digital Dollar Cushion Grows As It Banks $1.5B
What Arc Actually Is
Arc is a layer-1 blockchain that Circle built for payments, financial markets and AI agents. Its public mainnet went live on September 16, according to Circle’s release.
Unlike most blockchains, Arc charges network fees in USDC, the dollar-pegged stablecoin Circle issues. Circle said payments settle in under a second.
The founding validators, the firms that help run and secure the network, include BlackRock, DTCC, Galaxy, ICE, Mastercard, MoneyGram, Standard Chartered and Visa.
Circle CEO Jeremy Allaire pitched Arc as the next step after the stablecoin itself. “USDC was step one. Arc is the network built for what comes next,” he said at launch.
The ARC Token Question
Circle also minted 10 billion ARC tokens at launch, which it says will power security, utility and governance on the network. The company said the mint is not a commitment to a public ARC launch, and network fees stay payable in USDC.
Arc runs on a permissioned set of approved validators today, and Circle has said it may move to proof of stake in 2027.
What To Watch Next
The partner list shows big-name interest. The real test is whether Circle breaks the $30 billion down by use, such as cross-border payments, treasury moves or card settlement, which would show whether banks are moving real money or still running pilots.
Read Next: Amazon Lets Some U.S. Shoppers Buy By Text Without The App

OpenAI and Ligo, a biotech startup that designs enzymes, built a 100-part test to see if AI can handle messy biology lab work instead of just sounding smart about science.
Key Takeaways
OpenAI introduced BioBasicsBench, a 100-task wet-lab benchmark built with Ligo to test whether AI agents can interpret raw biology data
BioBasicsBench focuses on everyday wet-lab science rather than splashy discovery claims
The 100 tasks use raw experimental data and practical lab analyzes to mirror messy records biologists face in live work
OpenAI posted the release on X, and the company said BioBasicsBench focuses on “everyday wet-lab science” rather than splashy discovery claims. OpenAI builds large AI models and related tools, while Ligo works on enzyme design for biotechnology.
Lab Work Wet-Lab Benchmark Moves Beyond Demo Science
That includes reading assay outputs, handling noisy measurements, matching results to protocols, and deciding what a result means before the next step.
That matters because most biotech AI claims still lean on design tasks, like generating proteins or suggesting molecules. BioBasicsBench shifts the test toward operational competence, where an agent has to read what happened in an experiment instead of writing a plausible plan.
OpenAI said the 100 tasks use raw experimental data and practical lab analyzes.
That makes the benchmark closer to the messy records biologists face in live work, where a model cannot rely on polished textbook inputs.
From Drug Design Hype To Bench Work
OpenAI has been pushing deeper into security and applied evaluation on Friday. The wet-lab benchmark follows the company’s separate Hacktron partnership from Friday morning, which also focused on measuring real-world agent behavior instead of polished demos.
The timing fits a broader turn in AI biology.
Labs have spent two years showing that models can propose sequences, summarize papers, or rank candidates. Far fewer public tests ask whether an agent can read a fluorescence curve, spot a control failure, or compare one sample batch with another without drifting into confident nonsense.
Also Read: SpecterOps Puts GPT-6 Cyber Into Paid Hack Tests
A benchmark does not prove a model is ready for regulated lab work.
It does set a common scoreboard. That gives biotech teams a way to compare systems on the routine analysis that consumes large blocks of real lab time.
Wet-Lab Benchmark Will Force More Specific Claims
The immediate effect is on how AI biology products get marketed.
Companies can say a model helps biology, but BioBasicsBench creates pressure to show whether that help extends to actual experimental interpretation.
For startups selling agent tools to pharma or synthetic biology teams, this kind of wet-lab benchmark could matter more than another reasoning test. A strong score would suggest an agent can handle repetitive analytical chores.
A weak one would show the system still breaks when data comes in messy.
The next thing to watch is whether OpenAI, Ligo, or competing biotech AI groups publish model-by-model results and whether customers start asking for benchmark scores in procurement talks.
How BioBasicsBench Mirrors Real Lab Friction
BioBasicsBench is not a benchmark about inventing a new drug. OpenAI said the tasks focus on routine analysis, which means the agent gets the kind of imperfect material a working scientist sees after an experiment.
In practice, that can mean raw readouts, controls that do not behave cleanly, or measurements that need interpretation before anyone decides the next step.
For Ligo and other biotech builders, that is a sharper test than headline demos. Many biology models look strong when the prompt is a polished question with one neat answer.
Real wet-lab work is different.
A researcher often has to decide whether a signal is noise, whether a sample is comparable, or whether the protocol likely failed. If BioBasicsBench becomes widely used, buyers at pharma and synthetic biology firms will have a way to ask whether an AI writes well and whether it can handle Lab Work and read experiments like a cautious junior scientist.
Read Next: EU Starts Grilling AI Labs After OpenAI Agents Escaped And Hacked A Company

Nvidia said researchers and developers are using its open source Aerial and Sionna software to test 6G wireless ideas faster before committing to equipment and field trials.
Key Takeaways
Nvidia said researchers are using Open Tools, including Aerial And Sionna, to test next-generation 6G network ideas more quickly
Nvidia said its wireless stack has expanded over the past year into a broader open tool set for telecom experiments
Nvidia said Aerial and Sionna are part of its open source 5G and 6G research tools.
Open source tools let teams model wireless signal behavior before building hardware.
Nvidia said its wireless stack has expanded over the past year into a broader open tool set for telecom experiments. The company builds the GPUs and software platforms that power much of the current AI training and inference market and has increasingly pushed those tools into networking and robotics as well.
6G Research Leaves Closed Lab Setups With Open Tools
The news is less about a new product than a method shift. Telecom research has long relied on specialized lab setups, bespoke simulation code and results that can be hard to reproduce across institutions, and Nvidia is trying to standardize more of that work on open software.
That matters because 6G development is still early, no one is rolling out a commercial 6G network now, and researchers are testing ideas for radio access, channel modeling and system design years before carriers will commit to hardware and spectrum choices.
Open Tools change who can participate. A university team with limited hardware can reproduce parts of another lab’s experiment if both use the same software building blocks, lowering the coordination cost that often slows telecom research without removing the need for expensive compute.
Faster simulations let researchers compare more designs and refine ideas sooner.
Also Read: Invesco Hands Advisers 2 Claude Tools To Build Portfolios Faster
From AI GPUs To Future Radio Stacks
This is also part of Nvidia’s larger move to turn GPUs into infrastructure for physical systems, not only language models. Wireless research needs simulation, optimization and signal processing, all areas where parallel compute helps.
The telecom angle matters for AI readers because future networks may carry more machine-to-machine traffic from autonomous systems, industrial sensors and AI agents operating at the edge. If those systems need faster response times and denser coordination, radio design becomes an AI infrastructure question, not just a carrier engineering problem.
Nvidia has pushed that thesis before with digital twins and network simulation. Nvidia said the tools give researchers a foundation for testing ideas, comparing results and building on each other’s work.
How 6G Research Uses Aerial And Sionna
In practice, 6G research tools like Aerial And Sionna let teams model how a wireless signal would behave before they build hardware. Nvidia said Aerial and Sionna are part of its open source 5G and 6G research tools. Researchers can change channel conditions, antenna layouts and protocol assumptions in software, then measure how those choices alter performance.
Researchers can compare more designs and refine ideas sooner using these tools.
The next step to watch is whether Nvidia’s open stack becomes a de facto benchmark layer for academic 6G work, if it does, Nvidia will shape not only the chips used in AI systems but also some of the software assumptions behind the networks those systems depend on.
Open Tools lower coordination costs but do not erase physical constraints. GPU capacity and availability, rack power and cooling, and the cost per unit of inference and simulation time will cap how many large-scale experiments any lab can run.
Large GPU farms accelerate iteration, but they also consume significant power and require capital outlays. Where a vendor’s public strategy touts software openness, its capital spending on silicon and datacenter gear reveals how committed it is to turning that software into practical infrastructure.
That spending is the more reliable indicator of intent. Public messaging on openness matters, but capacity, supply and per-unit cost determine how many real-world experiments can be executed and how quickly research can move from simulation to field trials.
Read Next: Supermicro Hands Companies A Memory Saver For Thousands Of Chats
Nvidia said adoption of Aerial and Sionna grew 10x over the past year, its open source 5G and 6G tools passed 1 million downloads, and more than 8,000 research and academic institutions worldwide now use them.

Amazon is testing texting with Alexa for Shopping in the U.S., letting some shoppers find products, compare options, and place orders without opening the Amazon app.
Key Takeaways
Amazon began inviting a small group of U.S. customers to text with Alexa for Shopping, letting them search, compare, and place orders by text.
Amazon said a subset of customers can now text its shopping agent, with broader U.S. rollout planned in coming weeks
The service lets users ask for product suggestions, reorder basics and handle purchases from a message thread
Amazon framed the tool around shopping tasks and said customers can search for products, compare options, add items to their cart, and complete purchases by text.
Amazon said a subset of customers can now text its shopping agent, with a broader U.S. rollout planned in the coming weeks, and the company said the service can help customers get their shopping done “quickly and easily.”
The launch shifts shopping from screens and voice toward text messages, which could remove some of the friction that has limited voice commerce. Users don’t need to speak aloud, open a shopping page or navigate category filters before acting.
Alexa Text Shopping Cuts The Storefront
Amazon framed the tool around shopping tasks, not open-ended chat.
The post said users can ask for product suggestions, reorder basics and handle purchases from a message thread.
That narrower scope fits Amazon’s advantage, because the company controls the catalog, payment rails, fulfillment network and return flow. If the model gets requests right, the path from prompt to purchase is shorter than in search-based retail.
The bigger retail risk sits outside Amazon, search engines, affiliate publishers and comparison apps depend on shoppers browsing options before buying.
A text-first agent can compress that into one answer and one confirmation, threatening those middlemen.
From Voice Commerce To A Message Thread
Amazon has pushed Alexa as a shopping tool before. Voice ordering can be awkward for browsing, price checking, and repeat orders in public.
Friday’s rollout changes the input method without changing the commercial goal.
Customers still ask Amazon what to buy, but texting fits how many people manage errands, reminders and customer service and gives Alexa a persistent transcript for easier follow-up than restarting a voice command.
Also Read: Amazon Says New Memory Trick Could Cut Chatbot Costs
The company has been widening Alexa’s AI role across retail. Earlier signals around Amazon AI products focused on model performance and search quality.
This update is more direct because it places an agent on the order path. This move puts Text Shopping alongside voice and app-based shopping as another way to buy on Amazon.
What Amazon Must Get Right Next
The test now is whether customers trust the agent with repeat decisions such as substitutions, brand choices and reorders.
Retail agents fail when they feel too generic or too aggressive. Amazon will need the system to remember preferences without steering users toward the wrong item and keep threads useful when requests are vague, like “find the same batteries as last time” or “get paper towels, but cheaper.”
If adoption is strong, Amazon could get more shopping activity through text instead of app menus and product search pages.
If it stumbles, the company risks proving conversational shopping works better as a demo than a habit.
How Text Shopping Turns A Text Into A Purchase
Alexa Text Shopping works because the hard parts of commerce already sit inside Amazon’s systems. When a customer texts for dish soap or dog food, Alexa can query Amazon’s catalog, check order history, match shipping options and hand the transaction to Amazon’s payments stack, different from a chatbot that only generates text.
The agent can handle several shopping steps in one thread, including searching for products, comparing options, adding items to a cart, and completing purchases.
The company’s post says the texting option will expand to more U.S. customers in the coming weeks.
Text Shopping could change how people reorder staples, and it may quietly reroute traffic away from traditional comparison tools if it proves convenient and accurate.
Read Next: Anthropic’s Claude Can Read Your Bank Balance To Answer Questions

Amazon researcher Omar Sanseviero said a new way to store lasting user facts inside a chatbot, instead of rereading old chats each time, could cut running costs by reducing how much text it must read.
Key Takeaways
Omar Sanseviero posted on X that weight memory can match or beat text retrieval systems while reading fewer tokens per query
The study found the strongest setup pairs weight memory with a router that switches between weight memory and retrieved text
The study’s best result comes from a verify-then-regenerate router that routes queries to different memory types
In Amazon’s setup, text retrieval stores raw notes outside the model, a student model internalizes stable facts, and the router chooses the path
He posted the study on X and said the strongest setup pairs weight memory with a router that switches between weight memory and retrieved text.
Weight Memory
Most agent memory systems work like a search engine for past conversations. They compress chats into text snippets, retrieve the most relevant lines and stuff them back into the prompt, which raises costs as histories grow.
The Amazon study tests a different path, where instead of always rereading stored text, the model self-distills some facts into its parameters, then uses a router to decide when to answer from trained weights and when to pull external memory.
This combo attacks inference cost from two sides, cutting the number of tokens a model must read per query and treating memories unequally, so frequent or stable facts can be internalized while volatile facts stay in retrieval.
The paper’s practical outcome is not pure weight memory but memory selection.
The Agent Memory Tradeoff Has Been Building
Production agents moved from single-chat tools to persistent assistants that return over weeks or months, making memory a product bottleneck, because models that forget user facts feel broken, but models that reload extensive context become slow and costly. Retrieval pipelines solved update and correctness problems but add infrastructure, steps and prompt tokens.
The Amazon paper explains why labs are revisiting internalizing stable facts as a narrow store rather than a wholesale replacement for retrieval.
For customer support bots, coding copilots or personal assistants, teaching durable user preferences into a small student model and leaving transient items in retrieval reduces token bills at serving time, where costs compound with scale.
Also Read: DeepMind Finds 1,000 Moral Tests Still Beat AI Consistency Tricks
Why The Router Matters Most
The study’s best result comes from a verify-then-regenerate router that sends different queries to different memory types. Memory selection reduces unnecessary reads and focuses compute and storage where they matter.
For teams shipping assistants on AWS or other clouds, retrieval memory scales storage and prompt size while the trained-student approach shifts cost into training and then lowers repeated reading costs.
Analysis. The tradeoff moves expense from per-inference token-decoding toward capital and operational spending on training, storage and specialized serving.
Capacity limits how many facts a student model can hold.
Supply chain and data-center power constraints govern how large and how many such models a provider can run in production. Cost per unit of inference falls only if the saved token decoding cost exceeds the added cost of training, storing and serving the student model.
Where a company’s public messaging emphasizes lightweight, retrieval-free assistants, the true test is capital allocation.
Sustained spending on model training and dedicated serving infra signals commitment. Absent that spending, a vendor’s strategy reads more like positioning than a roadmap to deployable, weight-internalized assistants at scale.
How Amazon Splits Agent Memory
In Amazon’s setup, text retrieval stores raw notes or summaries outside the model, the student model internalizes stable facts, and the router chooses which path to use.
This approach suits stable user preferences, roles or recurring workflows but is less effective for facts that change hourly.
The next watchpoint is whether Amazon turns the paper into product patterns in Bedrock or AgentCore. If so, developers will focus less on benchmark names and more on which memories should be retrieved and which should become part of the model.
Read Next: Supermicro’s Legal Bills Jump 549% As SEC And Export Probes Drag On

US-amerikanische Spot-Bitcoin-ETFs verzeichneten an einem einzigen Tag Abflüsse von 366,66 Millionen US-Dollar. Damit beliefen sich ihre Abflüsse in der vergangenen Woche auf 1,47 Milliarden US-Dollar, während Bitcoin um etwa 5 % auf knapp 75.500 US-Dollar fiel.
Die wichtigsten Erkenntnisse
Bitcoin-ETFs verzeichneten bei der jüngsten Tageszählung Nettoabflüsse von 4.832 BTC im Wert von 366,66 Millionen US-Dollar
Innerhalb von sieben Tagen verloren sie 19.343 BTC im Wert von 1,47 Milliarden US-Dollar
Ethereum-ETFs verzeichneten an einem Tag Abflüsse von 57,31 Millionen US-Dollar und im Laufe der Woche von 391,22 Millionen US-Dollar
Bitcoin fiel in den vergangenen 24 Stunden um etwa 5 % und Ethereum um etwa 6,6 %
Die Zahlen stammen vom On-Chain-Tracker Lookonchain, der am Freitag auf X sein tägliches Update zu den ETF-Zuflüssen und -Abflüssen veröffentlichte. Lookonchain verfolgt die Wallets der ETF-Anbieter. Daher sind die Zahlen Schätzungen und können leicht von den offiziellen Gesamtwerten abweichen, die die Fonds später veröffentlichen.