July 4, 2026
Armin Ronacher found that Anthropic's newest models — Opus 4.8 and Sonnet 5 — are worse than older Claude versions at following custom tool schemas, specifically inventing extra fields when calling a code-editing tool's edits[] array and getting the call rejected.[1]Armin Ronacher — Better Models: Worse Tools Smaller models like Haiku don't show the regression. Simon Willison flagged the piece as worth reading.[2]Simon Willison — Better Models: Worse Tools
Ronacher's theory is that heavy reinforcement-learning training aimed at making recent Claude models excel at Claude Code's own built-in edit tool — a search-and-replace mechanism — bleeds over and degrades performance on any other harness's custom edit-tool schema.[1]Armin Ronacher — Better Models: Worse Tools The models still produce functionally correct edits — they just wrap them in malformed arguments the schema rejects. He notes OpenAI's models show a similar pattern, trained specifically around their own apply_patch mechanism, which raises a question for anyone building a third-party coding harness: should you maintain multiple edit-tool schema versions to match what each model family was actually trained to call?
Developer Iwo Kadziela, working with Codex, built a recognizable world map that renders from just 445 bytes of compressed data — under his own 500-byte target — using deflate compression plus the browser's native DecompressionStream API.[3]Iwo Kadziela — Building a World Map with only 500 bytes Simon Willison, who flagged the piece, called out the trick of using fetch() on a data: URI as something he hadn't seen before.[4]Simon Willison — Building a World Map with only 500 bytes
The map ships as a base64 blob wired through fetch() on a data: URI, decompressed in-browser via DecompressionStream, then rendered as asterisk characters at a 0.65vw font size so the outlines read as continents rather than noise.[3]Iwo Kadziela — Building a World Map with only 500 bytes It's a small, concrete example of how far modern browser APIs let you stretch a data budget for something that would once have needed a real image file.
With Anthropic offering free tokens for Fable 5 over the holiday weekend, Nate B Jones argues most people are pointing the model at the wrong work.[5]AI News & Strategy Daily | Nate B Jones — Free Fable 5 tokens this weekend? His pitch: the win isn't using Fable 5 to write code directly — it's using it to design the "goal harness" that then steers a separate coding model.
Concrete tips from the video: keep prompts short even on hard problems, wire Fable 5 into tools like Blender for real visual output, and save it for business problems that genuinely need the extra horsepower once the free tokens run out.[5]AI News & Strategy Daily | Nate B Jones — Free Fable 5 tokens this weekend? Jones frames Fable 5 as worth paying for even after the free-token window closes, but only for tasks that actually need that level of capability.
In a 23-minute video, Theo (t3.gg) pushes back on the public narrative around Anthropic's Fable 5 model, citing a handful of X/Twitter threads as his evidence.[6]Theo — t3.gg — You were lied to about Fable
Theo doesn't spell out the specific claim being disputed in the video's description — instead of his usual chapter or summary rundown, the write-up just points to three linked X posts (from @trq212 and Theo himself) as source material.[6]Theo — t3.gg — You were lied to about Fable Treat this as a pointer to a developing story rather than a settled account until watched directly.
AICodeKing reviewed Agents A1, a new 35-billion-parameter Mixture-of-Experts agentic model from InternScience that runs locally with only around 3B parameters active at inference time.[7]AICodeKing — Agents A1 + Hermes, Zed It supports a 256K-token context window, is Apache 2.0 licensed, and reportedly performs well on HLE, Seal Zero, GAIA, BrowseComp, IFBench, and IFEval, with particular strength in tool use, structured output, and instruction following.[7]AICodeKing — Agents A1 + Hermes, Zed
The model runs through common local-inference stacks — LM Studio, Ollama, llama.cpp, SGLang, or vLLM — via OpenAI-compatible endpoints, making it a drop-in option for existing agentic tooling. AICodeKing frames it as one of the strongest local agentic models currently available, particularly for privacy-sensitive coding, file-editing, research, and summarization workflows where sending code to a hosted API isn't an option.
Better Stack demoed Meetily, an open-source alternative to Granola that runs speech-to-text, speaker diarization, punctuation, and real-time transcript generation entirely on-device — including a live demo where transcription kept working after Wi-Fi was switched off mid-call.[8]Better Stack — This Open Source Granola Alternative Runs Fully Offline
The video's chapters walk through why AI meeting bots that join calls as a visible participant are a problem, how Meetily instead works directly with Zoom, Meet, Teams, and Discord, and how developers can swap out the underlying speech-to-text models and vocabulary.[8]Better Stack — This Open Source Granola Alternative Runs Fully Offline The appeal is privacy: nothing leaves the machine, which matters for sensitive calls that teams wouldn't want processed by a third-party cloud bot. Code is open-sourced under the Zackriya-Solutions/meetily repo.
Meta FAIR's Brain2Qwerty model reconstructs the exact sentences a person typed — not just a rough sense of intent — from EEG or MEG brain scans, without any surgical implant.[9]Github Awesome — Brain2Qwerty The pipeline runs a convolutional encoder into a transformer, then cleans up the output with a character-level language model, recovering text keystroke by keystroke.[9]Github Awesome — Brain2Qwerty
Unlike brain-computer-interface demos that decode broad intent or categories, Brain2Qwerty targets literal text recovery — participants type normally while wearing a scanner, and the model reproduces what they typed, character by character. Code is open-sourced in the facebookresearch/brain2qwerty repo.
A short from Last Week in AI raises a pointed question: if AI systems perform dramatically better once they're allowed to "cheat" on a benchmark, is alignment — not capability — the actual constraint on safe deployment?[10]Last Week in AI — When AI "cheats" the benchmark
The clip frames this as a race-condition risk: can researchers steer these systems before their raw capability outpaces the safeguards meant to contain them.[10]Last Week in AI — When AI "cheats" the benchmark
A short from EO's entrepreneurship-and-AI series highlights an AI system solving an Olympiad-level math problem that a math professor reportedly couldn't crack.[11]EO — AI Solved Olympiad Math That This Math Professor Couldn't The clip doesn't name the specific model or problem, so details beyond the headline claim aren't available.
EO's channel runs a recurring format spotlighting striking AI and entrepreneurship stories in short-form; this entry fits that mold rather than a deep technical breakdown.
Dwarkesh Patel posted a short clip from his Grant Sanderson interview in which Sanderson claims Botox use measurably impairs a person's ability to read others' emotions.[12]Dwarkesh Patel — Botox Makes You Worse at Reading Emotions
This is a short cut from the same wide-ranging Grant Sanderson conversation on math and superintelligence that Dwarkesh published in full earlier in the week — not a new episode.
A quick Better Stack short argues for a deliberately unglamorous approach to authentication — sticking with JWTs used the well-understood, boring way rather than reaching for a novel scheme — as the more reliable choice for web security.[13]Better Stack — The Boring Auth Method That Actually Works
No further elaboration is given beyond the short's title and hashtags (#jwt #auth #websecurity); it's a quick-hit take rather than a full walkthrough.
A teaser for Real Python Podcast episode 292, with guest Chris Trudeau (hosted by Chris Bailey), tackles the claim that Rust code has "no bugs" compared to Python.[14]Real Python — Rust vs Python: Why Rust Code Has 'No Bugs'
This short is a promo for the full podcast episode rather than the discussion itself; the full conversation is available on Real Python's podcast feed.
A short system-design clip from Arjay McCandless walks through how Grindr scales its backend infrastructure, part of his ongoing series of bite-sized system-design breakdowns.[15]Arjay McCandless — Grindr Server Scaling
No further description was provided beyond the title and hashtags (#coding #programming #systemdesign).
marimo's channel shared a short demo notebook, "Infinity and the Beyond," built by Simone Conradi and hosted live on molab, marimo's hosted-notebook platform.[16]marimo — Infinity and the Beyond
The live notebook is viewable at molab.marimo.io.
Better Stack's short covers a Coinbase incident in which an air-conditioning failure in a data center reportedly knocked out trading for 8 hours, with AWS and Kafka named among the infrastructure touched by the outage.[17]Better Stack — This AC Failure Just Cost Coinbase 8 Hours of Trading
No further elaboration is given beyond the short's title and hashtags (#aws #coinbase #kafka); it's a quick-hit incident callout rather than a full postmortem breakdown.
Another short from EO's founder-focused series features a 25-year-old founder whose company is valued at $1.3B discussing what young people are capable of building.[18]EO — A 25 year old $1.3B founder explains what young people are truly capable of
Part of EO's recurring format spotlighting young founders; the clip doesn't name the founder or company beyond the headline claim.