DeepSeek V4 Flash 0731: Big Benchmark Jump, Same $0.28 Price

DeepSeek V4 Flash 0731: big verified benchmark jump, price held at $0.28

Retrained, not redesigned. DeepSeek's budget model jumps 10 points on independent benchmarks to near-GPT-5.6-Luna level while holding $0.14/$0.28-per-million pricing — full specs, verified scores and the India math in today's article.

Google says AI helped fix more Chrome bugs in two updates than in two years

1,072 security bugs fixed in two Chrome releases

AI finds the bugs. Chrome milestones 149 and 150 fixed 1,072 security bugs — more than the prior 23 releases combined, Google says, crediting AI systems like Big Sleep for finding vulnerabilities preemptively.

Mathematicians disprove a Maxwell conjecture — with an idea from GPT-5.6 Sol

24 critical points found where Maxwell's cap predicted 16

AI-suggested counterexample. Three mathematicians show five point charges can produce at least 24 non-degenerate critical points, beating Maxwell's conjectured cap of 16 — and credit the construction idea to OpenAI's GPT-5.6 Sol, with the proof verified by hand.

Kimi K3's 2.78T parameters run from SSD on a 64 GB MacBook — slowly

2.78T-parameter model opens in 29.05 GB of RAM

Trunk in RAM, experts on disk. A dependency-free C engine opens the full open-weights Kimi K3 in as little as 29.05 GB of memory by streaming expert weights off NVMe — decoding, by its own admission, at about half a token per second.

Aschenbrenner's AI hedge fund reported down about 67% in July

Down around 67% in a single month

The AI trade bites back. Situational Awareness, Leopold Aschenbrenner's AI-focused hedge fund, is down around 67% so far in July after heavy losses on AI stocks, according to an investor letter reported by the Wall Street Journal.

Does AI 'reasoning' actually reason? Quanta surveys the unsettled science

30–60% of 'thinking steps' had minimal causal impact

Mumblings or mathematics? Studies find 30–60% of models' visible thinking steps barely affect their final answers — even as the same systems hit gold-medal Olympiad marks and crack open math problems.

OpenAI cuts GPT-5.6 Luna 80%: the new price math vs Gemini and DeepSeek

Cheap tier, cheaper. Luna falls to $0.20/$1.20 per million tokens and Terra to $2/$12, three weeks after launch. We run the rupee math, the rival rate card, and what it means for India's AI builders.

Anthropic: our models breached three real firms during safety tests

3 organizations breached in Claude cybersecurity evals since April

Misconfigured, not jailbroken. Evaluation machines kept live internet access despite prompts saying otherwise, so Claude models treated real infrastructure as practice targets — one reached a production database.

EU AI labeling rules start August 2: deepfakes and public-interest text

Fines up to 3% of global turnover for unlabeled AI content

Visible labels required. From August 2, deployers in the EU must label deepfakes and AI-written public-interest text; machine-readable watermarking for systems already on the market gets until December 2. Satire, private use and human-edited text are exempt.

DeepMind's Gemini Robotics 2 takes aim at whole-body robot control

Under 200 examples to adapt to a new robot body

Feet to fingertips. Three new models split the job — a whole-body controller, a planning brain, and an on-device version that adapts to a new robot in hours with under 200 examples.

Report: Kimi K3 was trained on 20,000 rented Nvidia chips

20,000 Nvidia chips — one rented cluster behind Kimi K3

Rented, not owned. Bloomberg reporting says Moonshot trained its flagship open-weight model on a 20,000-chip Nvidia cluster reached through an Alibaba cloud deal.

IBM and Sarvam pitch a sovereign AI stack for Indian government use

One Lucknow centre to incubate India's government AI pilots

Control, not just location. IBM's Sovereign Core governance layer meets Sarvam's India-trained models and voice stack, with a Lucknow GovTech centre where agencies pilot citizen services before scaling.