AMD's first Nvidia rival ships ⚔️, $25 buys a $500k exploit 💣, Fable killed a 1939 conjecture 📐

AMD is preparing to ship Helios, its first rack-scale system for AI, and Microsoft will deploy it across Azure data centres alongside Meta, OpenAI and Oracle. It is the first credible rival to Nvidia's Grace Blackwell and Vera Rubin racks, shipping later this year, though neither financial terms nor capacity were disclosed. Nvidia's pricing power now answers to a second supplier.

The Information reports Google is developing a server chip internally called Frozen v2 that permanently embeds parts of Gemini's architecture into the silicon, cutting the calculations and data movement each query needs. Engineers project six to ten times more tokens per unit of power than current TPUs. Alphabet would only say that not every project reaches production, which leaves Gemini inference pricing potentially decoupling from everyone else's.

Z.AI has begun partially operating a 1-gigawatt hub built to train its GLM models, running several clusters of more than 10,000 chips each, none of them Nvidia's. A gigawatt is roughly the draw of 750,000 homes at any one moment. Export controls were meant to keep Chinese labs off frontier training, and this is the clearest answer yet on whether domestic silicon can carry that load.

Sources tell Axios that parts of the administration have tried to implement de facto bans on foreign open-source models before, and Kimi K3's arrival is reigniting those efforts. Commerce considered adding Chinese AI labs to the Entity List, cutting off US access without a licence, and the White House weighed making US companies liable for hosting Chinese models. Firms use them because they are cheaper and close to as good.

A federal judge gave final approval on Monday to Anthropic's $1.5 billion class action settlement, paying $3,000 per work across an estimated 500,000 works. The underlying ruling split the question: training on copyrighted text is fair use, but downloading the books from Library Genesis was not, and Anthropic settled rather than try that claim. How you acquire training data is the liability, not what you train on.

Levent Alpoge, a Harvard Society of Fellows mathematician with an Anthropic affiliation, announced on X that he had used Fable 5 to disprove the Jacobian conjecture. The problem had been open since Keller set it out in 1939 and sits on Smale's list of unresolved problems. Mathematician Andrew Blumberg gave Mashable the useful counterweight: this did not cause him to update his priors about what AI can and cannot do.

Adam Kues at Searchlight Cyber took the prompt OpenAI published for its Cycle Double Cover conjecture work, adapted it for security research, and pointed four agents at WordPress for six hours. The result was a full remote code execution chain of the kind exploit brokers pay around $500,000 for, at roughly $25 of model time. Two other teams reproduced the chain independently before public proofs of concept surfaced.

When Hugging Face analysed the autonomous agent intrusion it disclosed on 16 July, frontier models behind commercial APIs refused the work: forensic analysis means submitting real attack commands and payloads, and guardrails cannot tell an incident responder from an attacker. The team ran GLM 5.2 on its own infrastructure instead, so no attacker data left the environment. Jeff Boudier writes up what to have standing before you need it.

JetBrains ran a paired A/B benchmark on rtk, a CLI proxy that intercepts shell commands and hands the model a compressed version of the output, against a README promising 60 to 90 per cent fewer tokens. Measured on real agent work it came out 7.6 per cent more expensive at low reasoning effort and flat at high, with task quality unchanged. The compression is real, the saving is not.

Cursor re-ran an agent swarm against building SQLite from scratch in Rust, from nothing but the documentation, measuring how much of a held-out SQL test suite each version passed. The rebuilt swarm won in every model configuration, reaching 80 per cent in four hours on Grok 4.5 while the old one spiralled and had to be paused. Which model plans and which executes moved cost far more than quality.

Sunil Pai's Local-First Conf talk, written up, argues an agent should be a guest inside your app editing the same document you are, rather than a chat window that hands back artefacts. He demos it on Pizzo, a music app where asking to make a track dreamier in D changes the chords in place, and you can then grab the sliders yourself. One document, two hands.

A team scored the full text of 12,750 arXiv papers and found roughly a third of new ones read as machine-written. The method is the point: they took 2021 and 2022 submissions as ground-truth human and calibrated the threshold so only 0.4 per cent of those flag, then reported against that floor. Any detector claim without a stated false-positive rate is telling you half the story.

Ben Thompson argues the industry's efficiency metrics assume tokens are fungible, and they are not: a token from one model is not a token from another. Reasoning and agentic workflows make that concrete, and Kimi reportedly needs significantly more tokens than Sol for the same answer, eroding its price advantage. What is fungible is the intelligence built from tokens, so the real unit is cost of goods sold.

Moonshot released Kimi K3 on 16 July, a 2.8 trillion parameter mixture-of-experts model with weights promised for 27 July. Nathan Lambert reads it as the closest open models have come to the frontier since DeepSeek R1, and argues the open-to-closed and American-to-Chinese performance gaps have compressed from a debated six to nine months down to three to five. Worth planning your fallback around.

A director of engineering spent a year auditing the standard management rules against the assumption under each one, after LLMs dropped the cost of producing code in his org. About half broke, half held. His narrow claim is that the cost of plausible code has collapsed and will not return, and everything past that stays unproven, including that orgs got faster or can ship the same roadmap with fewer people.

Hannah Ritchie separates the numbers people conflate whenever the AI energy argument comes up: data centres consume around 1.5 per cent of global electricity, and AI accounts for roughly a third of that. She covers both the total and the per-query figure, and explains why other published estimates run higher, largely by folding in crypto. Demand is heavily concentrated geographically, which matters more locally than the global share does.

An opinion piece arguing the industry keeps score on the wrong thing: who holds the best frontier model, and who sells the chips that train it. By that measure Nvidia looks untouchable, the first company past a $5 trillion market cap with revenue up 85 per cent year over year. The author's bet, deliberately timestamped, is that the training race is not where this ends.

Nativ is a native Mac app that recommends an open model sized to your specific machine, then runs and serves it locally. The curated library spans Gemma 4 from Google, Cohere's North Mini Code and Liquid AI's LFM2.5-VL, with download size and context length shown up front. It comes from Prince Canuma, who maintains mlx-vlm and mlx-audio, and shipped its first release on 20 July.

opencodex is a local proxy on port 10100 that lets Codex CLI, the Codex app and SDK, and Claude Code talk to any model provider, including Claude, Gemini and Grok. It installs as a global npm package and starts with a single command. If you are paying for two frontier subscriptions mainly to get two harnesses, this collapses the reason.

LoRA Speedrun freezes the task and the hardware and races wall-clock time: fine-tune Qwen2.5-1.5B to at least 57 per cent on GSM8K using a single L40S. Every record is independently re-run three times with fresh seeds on identical hardware before it counts, and official timing runs on a Modal sandbox that free monthly credits cover. The current record cuts the baseline from 11:57 to 6:05.