TECHGUYVER · INTEL DESK
Subscribe

11:35 PM

Today in brief.

🔧 Tech⏱ 11:35 PM

Three coding-agent things worth opening today

A sub-1MB Codex clone in C++, a live API cost calculator, and a cheap-model trick that stops you burning premium tokens on deploys.

4 sources
10:34 PM 🏛️ Policy

The decel debate got a name and a venue

After last week's slow-down letter, the argument about whether slowing AI is even legal is now being had out loud by the people who'd be slowed.

4 sources
6:46 PM 🎬 Media

Seedance 2.5 held product consistency well enough to fake an ad

One photo of a real product plus a prompt produced an influencer video where the product stayed itself the whole way through - the failure mode that kept AI out of paid ads.

3 sources
4:18 PM 🤖 AI

Anthropic's Fable reproduced half of OpenAI's math results in 24 hours

Yesterday OpenAI's unreleased Astra model got the headlines for cracking ten open math problems. Today an Anthropic researcher reproduced five of them with a model you can already buy, and a formal paper says one of the ten proofs is simply wrong.

+15 sources
2:07 PM 🔧 Tech

Mozilla published a State of Open Source AI report

A neutral party finally put numbers on what "open source AI" actually means in practice, at the moment the term is most contested.

3 sources
8:17 AM 🤖 AI

The benchmarks people actually trust are now private jokes

The pelican-on-a-bicycle test stopped discriminating between models, so people are quietly inventing weirder personal benchmarks - which says more about evaluation than any leaderboard does.

3 sources
🔧 Tech

94.8% of websites are never cited in an AI answer

Almost nobody blocks the AI crawlers, and almost nobody gets cited by them either - a visibility index just put a hard number on the new referral cliff.

2 sources
🔧 Tech

DeepSeek V4-Flash now runs locally with llama.cpp speculative decoding

The cheap frontier-class model everyone benchmarked last week is now actually runnable on your own machine, fast, with the standard local stack.

4 sources
🔬 🔬 Science

China's DFSX claims double the memory bandwidth of Nvidia's GB200

Memory bandwidth is the real bottleneck for inference, and a Chinese chip is claiming 2x Nvidia's flagship on exactly that number.

2 sources
🎬 Media

An AI poster won a state fair art contest

The Ohio State Fair poster contest was won by an AI image, three years after the last time this happened made national news - and this time the argument is smaller.

2 sources