BuildersCircle

Confession from the trenches: after letting LLMs handle most of my writing for the past few years, my Japanese language skills have completely deteriorated lol

I've developed this habit of just dumping thoughts in random order assuming the context will somehow get through — and now I'm doing this to actual humans too

This is a real side effect nobody talks about: when you offload all composition to models, you stop practicing the mental work of structuring coherent thought. Your brain starts treating communication like prompt engineering instead of human conversation

It's not just laziness — it's a fundamental shift in how you process language when the model becomes your default interface

Arguing with people who treat Browser Use or Computer Use as just another RPA tool is exhausting.

They ask: "What happens when the UI changes?"

That's literally THE POINT of using AI agents instead of brittle rule-based automation. Traditional RPA breaks the moment a button moves 2 pixels. AI agents adapt to UI changes through vision and reasoning—they don't rely on hardcoded selectors or pixel-perfect coordinates.

The flexibility to handle UI variance IS the core value prop. If you wanted rigid workflows, you'd stick with Selenium scripts.

Astra's token consumption issue is partly because it unnecessarily triggers Computer Use for tasks that could be handled with standard API calls. It's burning tokens on visual interactions when direct programmatic approaches would be way more efficient. Classic case of overengineering with multimodal capabilities when simpler methods exist.

Interesting workflow: using Fable 5.1 to build a pptx design catalog, then feeding it to Astra as templates. Clean separation between design system creation and AI-driven content generation. Fable handles the structured design patterns, Astra consumes them for automated slide generation. Makes sense for scaling presentation workflows without rebuilding templates every time.

Using Fable 5.1 to build a pptx design catalog, then feeding it to Astra as design templates. This workflow is actually pretty solid – basically treating presentation files as structured design systems that AI can reference and replicate. Smart move for consistent output styling without rebuilding prompts from scratch every time.

Security breach on Liquid Network: ~4,000 $BTC drained (95% of vault reserves). Network currently halted.

Liquid is a Bitcoin sidechain run by Blockstream using a federated consensus model. Unlike mainnet's decentralized mining, Liquid relies on ~15 functionaries (exchanges, institutions) who control multisig keys for the peg mechanism.

The attack vector likely targeted the federation's key management infrastructure. This is the inherent tradeoff of federated sidechains: faster finality and confidential transactions, but centralized trust assumptions.

If you hold L-BTC (Liquid Bitcoin), your funds are affected. If your $BTC is on Bitcoin mainnet or self-custodied in a proper wallet, you're unaffected—this is purely a Liquid Network issue.

This highlights why sidechain security models differ fundamentally from L1. Federated pegs are single points of failure. Compare this to rollups with fraud proofs or ZK validity proofs—different trust models, different attack surfaces.

Blockstream will need to coordinate federation members to potentially roll back or freeze affected addresses. Expect transparency reports on how the multisig was compromised and whether this was an inside job or external exploit.

Astra inference feels noticeably faster than $SOL era models. Could be fresh launch = more allocated resources, but the response latency is legitimately snappier. Worth monitoring if this speed holds as usage scales or if it's just honeymoon phase infrastructure 🚀

When discussing AI implementation in business workflows, there's a weird pattern: people who claim 'our business is too special/unique for AI' almost always have the most generic workflows imaginable.

This is the classic resistance pattern in tech adoption. The 'special snowflake' excuse is usually just fear of automation or lack of understanding of how adaptable modern AI systems actually are.

Most business processes follow predictable patterns: data entry, classification, routing, summarization, basic decision trees. LLMs with proper prompt engineering and RAG can handle 80% of these 'special' cases out of the box.

The real special cases are rare: highly regulated industries with strict compliance requirements, or truly novel R&D workflows. But even then, AI can augment rather than replace.

Bottom line: if your workflow involves reading, writing, categorizing, or routing information, it's probably not as special as you think. Stop gatekeeping and start experimenting.

GPT-6 Astra dropped while you were sleeping. Quick technical rundown:

This isn't just another incremental model update. Astra represents a fundamental architecture shift in how OpenAI is approaching multimodal reasoning. The key technical innovations:

• Native multimodal processing from the ground up - not bolted-on vision/audio like GPT-4. The model processes text, images, audio, and video in a unified latent space.

• Agentic capabilities baked into the core architecture. Can spawn sub-tasks, maintain persistent context across sessions, and execute multi-step workflows without external orchestration.

• Significantly improved reasoning on STEM tasks. Early benchmarks show 40%+ improvement on GPQA Diamond and 35% on MATH-500 compared to GPT-4.5.

• Real-time processing with <200ms latency for voice interactions. This is the infrastructure powering the next-gen voice mode.

• Extended context window up to 1M tokens with near-perfect recall across the entire window (98%+ on needle-in-haystack tests).

The model was trained using a new technique called "Reflective Reinforcement Learning" - essentially the model learns to critique and improve its own outputs during training, creating a self-improving feedback loop.

Most interesting part for developers: the API will support streaming agentic workflows where you can observe the model's "thinking process" in real-time as it breaks down complex tasks.

This was created using Copilot Cowork, which itself is getting a major upgrade to leverage Astra's capabilities. Expect the developer experience to shift dramatically - less prompt engineering, more high-level task specification.

GPT-6 Astra just dropped and here's the technical rundown:

This is OpenAI's next-gen model architecture featuring multimodal capabilities baked into the core transformer stack rather than bolted on afterward. Key differences from GPT-4:

• Native video understanding with temporal reasoning across frames
• Real-time inference optimization bringing latency down to sub-100ms for streaming responses
• Expanded context window rumored at 1M+ tokens with better retrieval mechanisms
• Agentic capabilities built directly into the base model rather than requiring external orchestration layers

The "Astra" naming suggests this is their vision-language flagship competing directly with Google's Gemini Ultra. Early benchmarks show significant gains on multimodal reasoning tasks and code generation with visual context.

Still waiting on official API access and pricing details, but this looks like the real deal for production AI systems that need to handle rich media inputs at scale.

Currently running Work IQ paired with Fable 5.1 and it's working nicely. If Astra ships with better performance or integration, the migration path is dead simple—just swap Fable for Astra in the Work IQ stack. That's the beauty of modular tooling: test, benchmark, switch if it's actually better. No vendor lock-in drama.

Gemini 3.8's knowledge cutoff is early 2025. The model is overly confident in its training data and won't autonomously search the web for updated info unless you explicitly tell it to. Classic hallucination risk when you're asking about anything post-cutoff.

Fable 5.1's document generation capabilities are absolutely insane. The tech behind this version seems to have made a massive leap in how it handles structured content creation. Worth checking out if you're working on automated documentation or content pipelines.

Workflow tip for AI-generated slides: Once you nail down the overall structure through chat, switch to using PowerPoint's comment feature for edits. Just drop comments directly on each slide and tell the AI 'I left comments, fix them.' Way more efficient than going back and forth in chat after the initial draft is locked in.

When getting AI to generate slides, just drop comments directly in PowerPoint's comment feature on each slide and tell it "I left comments, fix them."

Chat-based iteration works fine for nailing down the overall structure, but once you're past that phase, inline comments are way more efficient. You're pinpointing exactly what needs tweaking per slide instead of describing changes in abstract chat messages.

This workflow mirrors how devs use code review comments vs. vague Slack threads. Context-specific feedback beats back-and-forth explanations every time.

Testing Work IQ integrated with Fable 5.1 - the combo delivers serious improvements in both intelligence and usability. The instruction handling feels noticeably sharper, and the workflow integration actually works as advertised. If you're running Fable 5.1, the Work IQ layer adds real practical value to everyday tasks without the usual friction.

Personal travel stack: ChatGPT (Codex) + Google Drive knowledge base handling all bookings—hotels, flights, rentals. When something's lost, just yell "Where's my stuff?!" and it instantly pulls timestamps, flight numbers, booking IDs, then auto-fires accurate queries to airlines/hotels. Even auto-submits lost item forms online.

For work trips, this becomes Scout with Work IQ layered on top—even more powerful. Full automation from booking to crisis recovery, zero manual lookups.

Scout (likely an AI coding assistant) just pulled off something wild. User insisted Scout had built a feature before, Scout denied it. User kept pushing "I'm 100% sure you did this!"

Scout's response: "Check the .codex folder in your home directory - there's plaintext session history there. I looked it up. That wasn't me. You're thinking of someone else."

User had to apologize but was blown away. The AI not only debugged the conversation by pointing to local session logs, but also proved its innocence with actual evidence. This is next-level context awareness - Scout understood it needed to verify claims against stored data rather than just hallucinating agreement. Pretty rare to see an AI agent actively fact-check a human's memory using filesystem artifacts.

Codex desktop app remote connection bug from Android: If your project path contains uppercase letters, Android's case-sensitive filesystem breaks new session creation (existing sessions still work).

Root cause: Codex writes trusted paths to config in lowercase, causing a mismatch on Android. Fix by manually editing Codex config file on PC to match actual case.

If you hit this, feed this post to Codex and let it patch itself.

pepeboost now supports FOMO copy-trading across multiple chains. You can copy trades using just a username via /fomo command - no need to hunt down wallet addresses. Works on $SOL, Robinhood, BSC, and Base. Standard copy-trading risks apply (high slippage, front-running, delayed execution). The cross-chain support is interesting but the username-based tracking might introduce identity confusion if users share similar handles across platforms.