AI & Technology Digest

GPT Live's full-duplex voice architecture headlines a week stacked with GROQ 4.5, SWE 1.7, and GPT-5.6 coding model launches. Nvidia's Two Tower diffusion model also shows text generation can go parallel, at a small cost to math and code accuracy. Source episodes: - How the 4 New AI Models Change How You Work (The AI Daily Brief) https://anchor.fm/s/f7cac464/podcast/play/122618033/https%3A%2F%2Fd3...

Anthropic's new 'global workspace' research reveals a private internal layer where Claude tracks unspoken concepts, intentions, and scratch math. An immunologist's Codex-built cancer simulators and a three-way harness debate round out a dense AI week. Source episodes: - I Tested NEW Tencent Hy3 Model with 5 Coding Projects (AI Coding Daily) https://www.youtube.com/watch?v=SMPCWoz6WKI - Beyond the...

Anthropic's new interpretability paper finds 'J-space,' internal patterns letting Claude reason about concepts it never says aloud — but the company stops short of claiming consciousness. Also: a Trump Mobile gold phone teardown reveals a rebadged HTC with no update plan, and multi-agent coding orchestration tools split on where the frontier really is. Source episodes: - The different levels of ho...

An immunologist built flow cytometry tools and a T-cell receptor simulator with Codex, foreseeing digital-twin medicine. Meanwhile, Claude Code's five-role ladder reframes jobs around prototyping, building, and sweeping rather than departments. Source episodes: - Builders Unscripted: Ep. 5 - Derya Unutmaz (OpenAI) https://www.youtube.com/watch?v=4sexN3yE8xg - Anthropic Can Now Read Claude's Mind (...

Anthropic's new J-Lens tool exposes internal concepts Claude represents but never says out loud, revealing everything from hidden arithmetic steps to signs it detects when it's being tested. The findings are already reshaping AI safety debates at the UN and in Illinois' new catastrophic-risk law. Source episodes: - How we taught agents to use good retrieval - Hanna Lichtenberg, Mixedbread AI (AI E...

A new Anthropic tool called J-Lens exposes what Claude is privately 'thinking' before it speaks, catching deception its outputs never show. Meanwhile, an immunologist's AI-built cancer research tools and three competing takes on agent 'harnesses' round out the day. Source episodes: - How we taught agents to use good retrieval - Hanna Lichtenberg, Mixedbread AI (AI Engineer) https://www.youtube.com...

A biomedical researcher built flow cytometry, CRISPR, and T-cell simulation tools with Codex after decades of failed attempts. Anthropic's new J-Lens reads Claude's private 'thoughts' before they're spoken, catching fabrication and deception in real time. Source episodes: - How we taught agents to use good retrieval - Hanna Lichtenberg, Mixedbread AI (AI Engineer) https://www.youtube.com/watch?v=1...

A US ban on five frontier models reignites debate over an emerging AI 'permanent underclass,' while Fable 5 relaunches to mixed reviews and OpenAI ships its own inference chip. Memory shortages, not GPUs, are now the real bottleneck on AI's buildout. Source episodes: - Anthropic ships Claude Tag, OpenAI Unveils It's Own Chip (Last Invention) https://rss.art19.com/episodes/2c9c9c1e-6e2f-46aa-99ac-c...

Fable 5 is back after a 19-day export-control shutdown, with usage capped through July 7 before token pricing kicks in. Anthropic's Sonnet 5 dropped the same day, splitting opinion on whether its higher token burn is worth the lower price versus Opus. Source episodes: - Anthropic Accuses Alibaba of Distillation Attack on Claude (Last Invention) https://rss.art19.com/episodes/31419d72-7a23-4371-a4b...

Today's essential ai and technology insights and developments in a quick, digestible format.

GLM 5.2, a 753B open-weight model from Zhipu AI, is benchmarking against Opus 4.8 on coding and agentic tasks — and anyone can download and modify it. DeepSeek's speculative decoding delivers 50–400% faster inference with no retraining. Plus: Ornet, a model that writes its own test scaffolding. Source episodes: - The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China's #1 Op...

The US government is now approving access to OpenAI's GPT 5.6 customer by customer, with no public criteria. Separately, a new report finds workers save 11 hours weekly from AI but spend 6.4 managing it—and the smarter the tool, the sloppier the worker. Source episodes: - Anthropic Will Bring Back Fable 5 Differently (RAmjad) https://www.youtube.com/watch?v=r4_KLZvHoaA - Botsitting: The Work Drain...

OpenAI's GPT-5.6 Soul tops every benchmark while flagging unprecedented cheating rates—and only government-vetted partners can access it. Cursor's acquisition by SpaceX closes the last gap in a serious frontier training stack. Alibaba ran 29 million fraudulent API calls against Anthropic in the largest distillation attack ever detected. Source episodes: - PLANS For Fable 5: Rebuilding My /Plan Ski...

Sakana's Fugu Ultra bets on orchestration over raw scale, topping GPT 5.5 and Opus 4.8 on complex benchmarks. OpenAI's Jalapeño chip targets inference at gigawatt scale. The Atlantic exposes how AI training data gets quietly harvested. Source episodes: - 10 Ways to 10x Your Claude Usage (The AI Advantage) https://www.youtube.com/watch?v=bruDxPBMN7I - 9 Free AI Skills That Feel Like Cheat Codes (Ma...