Anthropic says its frontier model autonomously produced two publishable cryptanalysis results, cutting HAWK-256's effective security and speeding an AES attack 200-800x.
The first verifiable demonstration that a frontier model can produce publishable cryptanalysis, with a benchmark released so anyone can check the claim.
⚡ Quick hits
AgentsFireworks AI launched Nexus, a drop-in router that sends routine coding requests to open-weight models, reporting 3–5x cost cuts and 33% lower cost per merged PR.MarkTechPost ↗
A letter signed by researchers at Anthropic, OpenAI, Google and Meta asks the US to lead an international effort to build tools for slowing frontier AI.
⚡ Quick hits
AIGCIdeogram said it is partnering with AMD to serve its open-weight Ideogram 4.0 image model at scale on Instinct accelerators.X @ideogram_ai ↗
ModelsLMArena said xAI's Grok 4.6 arrives August 7 and will enter the arena leaderboards next week.X @arena ↗
ResearchOpenAI published a report on agentic scientific computing, using eight industry and academic case studies to argue coding agents shift researchers toward verification and stewardship.OpenAI ↗
Open SourceAi2 detailed its OlmoEarth Platform, which ran a North America wildfire-risk job on 19,600 CPUs and 994 GPUs for a 155x speedup over serial inference.Hugging Face ↗
ResearchLMArena's Code Arena added fullstack web development rankings, scoring models on multi-step reasoning and tool use across complete app-building tasks.X @arena ↗
Business Insider reports Amazon has halted development on Nova Premier, Omni, Reel and Canvas, shifting resources to a research group under Pieter Abbeel.
Google's hosted agent runtime switches its default to Gemini 3.6 Flash and adds pre/post tool-call hooks, token ceilings, cron triggers and a free tier.
⚡ Quick hits
PolicyPJM, the largest US grid, will curtail power to data centers of 50MW or more during shortages from June 2027, with advance notice and compensation.TechCrunch ↗
AgentsElevenLabs extended ElevenAgents to SMS, Telegram, email and ticketing, with per-channel response controls so one agent keeps the same knowledge, tools and guardrails everywhere.X @ElevenLabs ↗
Open SourceLiquid AI open-weighted LFM2.5-Encoders, 230M and 350M encoders with 8,192-token context that run about 3.7x faster than ModernBERT-base on CPU.Hugging Face ↗
AIGCRunway announced that ByteDance's Seedance 2.5 video model is coming soon to its platform, adding a rival lab's generator to Runway's tool lineup.X @runwayml ↗
FundingFish Audio raised a $52M seed led by Coreline Ventures and Capital Today for voice models with 15,000+ controls, citing 8M users and $21M ARR.TechCrunch ↗
IndustryNvidia is investing in Ilya Sutskever's Safe Superintelligence, moving the ~$30B-valued lab off Google TPUs onto Vera Rubin GPUs for a claimed tenfold compute boost.The Decoder ↗
Alibaba Cloud says its Zhenwu M890 supernode is the first Chinese system to run Moonshot's 2.8-trillion-parameter model; Huawei Ascend also claims day-0 support.
⚡ Quick hits
AIGCRecraft shipped V4.1, showcasing jewelry renders with physically convincing glass refraction and depth across rings, earrings and collars.X @recraftai ↗
A reported 81.5% of Samsung foundry staff want out within two years as rival SK Hynix pays bonuses more than three times larger on the back of AI memory profits.
Cursor's first localised plan bundles Composer 2.5, Grok 4.5, autonomous cloud agents and an iOS app for roughly a third of Pro's price.
⚡ Quick hits
ResearchApple detailed the audio detokenizer behind Siri Expressive Voices: decoupled temporal-depth diffusion streaming at ~21MB peak memory, 16x faster than real time, +0.28 MOS.Apple ML Research ↗
ResearchAn engineering writeup claims a $500 reinforcement-learning fine-tune of a 9B open model beat frontier models on a product-catalog review task.Fermisense ↗