Thought this was a joke, but i double checked. US Gov Federal Register use distilled Qwen models in their search mode. The irony.

‼️ BREAKING: Chinese labs DeepSeek and Moonshot were quietly relaying customer prompts to Claude through fraudulent accounts. Users thought they were using Chinese AI, but were actually getting answers from Claude. Their customers' sensitive data now is in Anthropic's hands. One user, who is tied to China's military, asked Kimi whether a person tracked across hundreds of CCTV cameras in Chengdu was behaving abnormally. DeepSeek passed along live credentials for a Russian government database, shared by an IT operator handling data from a Defense Ministry-linked agency. The two Chinese labs relayed customer prompts to Claude through fraudulent accounts and saved the exchanges to train their own models. Moonshot forwarded almost 300,000 customer requests in a single ten-day stretch. DeepSeek flagged users running its models inside coding tools like Claude Code and OpenCode, then routed some of them to Opus.

Sam Altman on Bloomberg Television: GPT-6 Astra is not the model that was paused over cybersecurity concerns; the paused model is a future model. "Astra has been done training for a while. The model that we recently talked about pausing, uh, is a future model. But with Astra, we did hit cyber-critical." ---- From "Bloomberg Television" YouTube channel, (full video link in comment)

GPT-6 Astra by @OpenAI achieves SOTA on ARC-AGI: - Astra scores 63% on ARC-AGI-3, 99% via a new provider adapter harness - It surpasses human performance on 96% of ARC-AGI-3 levels - It builds the most precise symbolic model of novel environments we've seen Our analysis:

Muse Spark 1.3 is rolling out today with frontier performance almost too cheap to meter. This is the biggest jump we've made so far on coding and agentic work. Try it in Muse Code and our API. Next up 🍉 and Muse Spark open weights releases coming soon.

NVIDIA and Hugging Face have had serious acquisition conversations in recent weeks about a deal that would value Hugging Face at over $13 billion.

Holy moly. 🤯 Qwen 3.8 27B just scored higher than: GPT 5.6 Terra GLM 5.2 DeepSeek V4 Pro Muse Spark 1.2 Claude Opus 4.8 on the Artificial Analysis Agentic Index. And you can run this on a single RTX 3090/4090. Go show your GPU some respect. 🫡

So Anthropic scraped the entire internet to train Claude, yet somehow completely omitted the parts where the CEO's wife ran a porn startup, married a man over 40 years older and pitched her app to Jeffrey Epstein. Peak AI alignment right there. xitter.cf/amir/status/2088014229…

DeepSeek V4 Pro 0813 scores 53 on the Artificial Analysis Intelligence Index, 8 points above April's DeepSeek V4 Pro - but with a 3.6x price increase and only 1 point above DeepSeek V4 Flash 0731 @deepseek_ai has released DeepSeek V4 Pro 0813, its new flagship model, along with updated pricing. With the new pricing, it still sits on our Pareto frontier for Intelligence vs. Cost, but by a smaller margin than previous DeepSeek releases. Under the new pricing, DeepSeek’s first-party API will charge $1.32 / 1M input tokens and $3.96 / 1M output tokens - an increase of +264% on the blended price. Cached input tokens receive a ~97% discount (rather than the previous 99%): cache hits are priced at $0.044 / 1M tokens, 12x higher than previous cache-hit pricing. Off-peak pricing is discounted by 50%. This new pricing takes effect on August 16, 2026 - until then DeepSeek is offering V4 Pro 0813 at the same pricing as the earlier V4 Pro model. DeepSeek published the model weights under the MIT license, making it the second most intelligent open weights model we have benchmarked; however, weights for Qwen3.8 2.4T A95B have recently been released and will be evaluated on the Intelligence Index soon. DeepSeek V4 Pro 0813 remains 7 points behind Kimi K3, the current open weights leader. DeepSeek V4 Pro 0813 retains the previous version’s architecture at 1.6T total parameters and 49B active parameters, with a context window of 1M tokens. Most of DeepSeek V4 Pro 0813’s gains over the April version are in agentic capabilities. DeepSeek V4 Pro 0813 is more token efficient than its predecessor, generating ~30% fewer total output tokens across Artificial Analysis Intelligence Index. Key results: ➤ DeepSeek V4 Pro 0813 is only 1 point ahead of DeepSeek V4 Flash 0731 on the Artificial Analysis Intelligence Index, and has ~3.8x the active parameters. The two models tie on Terminal-Bench 2.1 at 79%. For agentic and coding benchmarks the smaller model performed at around the same level. ➤ DeepSeek V4 Pro 0813 is more token efficient than DeepSeek V4 Pro, using 128M output tokens to run the Intelligence Index, ~30% fewer than the April version. This places DeepSeek V4 Pro 0813 on the open weights Pareto frontier for intelligence versus output tokens, with fewer tokens used than Kimi K3 (133M) and GLM-5.2 (141M). However, it lags behind many frontier models in its intelligence tier. GPT-5.6 Sol scores 61 on the Intelligence Index but only uses 70M tokens in total. ➤ Cost per Task increases significantly despite an improvement in token usage, driven by DeepSeek’s price increase. At $0.25, DeepSeek V4 Pro 0813’s Cost per Task is 5x that of DeepSeek V4 Pro ($0.05), but ~20% lower than GLM-5.2 ($0.32 Cost per Task). Despite the increase in price, DeepSeek V4 Pro 0813 still sits on the Intelligence vs. Cost per Task Pareto frontier, but barely - it is just one cent below Gemini 3.7 Flash (medium) on Cost per Task. ➤ DeepSeek V4 Pro 0813 makes significant improvements in real-world agentic knowledge work. On GDPval-AA v2 its Elo improved to 1590, +284 points from DeepSeek V4 Pro (1306), which implies an ~84% expected win rate against the April release. Its GDPval-AA v2 Elo is ahead of earlier-generation proprietary models like Claude Opus 4.7 and GPT-5.5, but it trails Kimi K3. DeepSeek V4 Pro 0813 works for longer on GDPval tasks than before, averaging 39 turns per GDPval-AA task compared to 23 for the April model. ➤ DeepSeek V4 Pro 0813 makes a 12 point improvement in AA-Omniscience, driven entirely by increases in accuracy. DeepSeek V4 Pro 0813 scores +1, up from -11 for DeepSeek V4 Pro. Its AA-Omniscience accuracy rises 6 points from 43% to 49%, while regressing slightly in hallucination rate from 94% to 95%. Compared to peer models, DeepSeek V4 Pro 0813 has a high attempt rate of ~99%, meaning it almost never declines to answer. By contrast, Kimi K3 attempts only 77% of questions and has a 53% hallucination rate. Additional model details: ➤ Context window: 1M tokens ➤ Weights: 1.6T total parameters; 49B active parameters ➤ License: MIT ➤ Accessibility: Accessible through DeepSeek first-party API at launch ➤ Pricing: On DeepSeek’s first-party API, updated pricing will take effect on August 16 at $1.32 / 1M input tokens and $3.96 / 1M output tokens. Cached input tokens receive a 97% discount, priced at $0.044 / 1M tokens. Off-peak pricing is discounted by 50%.

Alongside the release of GLM-5.3, we are launching the OpenVuln project: huggingface.co/spaces/zai-or… If you own or maintain an open-source project and would like additional security support, you can submit its public GitHub repository. OpenVuln will scan the repository for potential vulnerabilities and help you safeguard your project.

BRO WHY IS ANTHROPIC SO QUIET. suspiciously quiet.

🚨 JUST IN: Claude models will now have invisible watermarks embedded in ALL text, and ALL metadata attached to files…

Absolutely insane. This might be the clearest glimpse yet of how AI will transform scientific discovery. Anthropic asked an unreleased version of Claude to take a real stab at the Riemann Hypothesis, one of the most famous unsolved problems in mathematics. It failed. But while failing, Claude unexpectedly improved a longstanding lower bound for the proportion of zeros of the Riemann zeta function known to satisfy the hypothesis, from 41.6% to 67.2%. Claude coordinated 60 subagents, tested hundreds of ideas, searched the literature, challenged its own proof, and formalized the result in Lean. This is what the beginning of autonomous AI-driven scientific discovery looks like. Unironically, the Rienann hypothesis will probably be solved in the very near future. I am speechless.

🚨 Composer 3 Leaks: It will drop > It outperform Opus 5 and GPT-5.6 Sol on coding and agentic tasks > Six different internal variants have reportedly appeared > Reasoning levels range from Fast → Medium → High → XHigh > It Will be around 5x cheaper than opus 5 and GPT-5.6 Sol > Public release could arrive Early This month Is Composer 3 about to become the new king of AI coding agents?