Stanford researchers and collaborators introduce Germinal, a generative pipeline that designs antibodies for a chosen protein epitope by co-optimizing AlphaFold-Multimer structure scores with an antibody language model. In Nature Biotechnology, they report testing only 43 to 101 designs per antigen across four targets, including PD-L1 and a viral protein, with hit rates of roughly 4% to 22% and some nanobodies reaching nanomolar affinities after mammalian expression. The claim is de novo binders without huge library screens, a faster path to molecular tools and drug leads. Caveat: wet-lab filtering remains essential, and a binding hit is not yet a developable therapeutic.
Sydney AI infrastructure firm Firmus said it secured full commitments for a $2 billion strategic equity round, with follow-ons from Coatue and Nvidia plus new money from Blackstone vehicles and Jane Street. Capital will accelerate Project Southgate AI factories across Australia and prepare Asia-Pacific expansion, including early Indonesia work, on Nvidia's DSX architecture. Firmus says new equity over the past year now exceeds $3 billion, with post-money valuation above $10.5 billion. The bet is vertically integrated capacity that turns energy into tokens via HyperCube hardware and grid-aware software. Caveat: figures are company stated, and delivery still depends on power, permits, and GPU demand.
Reuters reports Alibaba plans to ask large commercial users of its next open-weight Qwen model for a share of revenue those deployments generate, according to two people familiar with the plan. The terms would arrive with the expected Qwen3.8-Max weight drop, following Moonshot's Kimi K3 license that pushes service providers above about $20 million in annual sales into commercial talks. Alibaba already charges for cloud-hosted Qwen but has largely left self-hosted open weights unpaid. Chinese labs appear to be testing freemium licensing as open models chase U.S. rivals. Caveat: the rate is unconfirmed and rests on anonymous sources.
Kily raised Rs 30 crore led by Sorin Investments, with Razorpay and Wyser Capital joining, Economic Times reports. The 2025 startup, founded by former Flipkart, Amazon, and Affle operators, builds agents that price, advertise, and restock for brands on Amazon, Flipkart, Blinkit, Zepto, and Instamart. It says ITC is already a client, and proceeds will deepen product work and sales to large consumer brands. The pitch is simple: marketplaces already optimize with AI, so brands need software that executes, not another dashboard. Caveat: the round is early, with little public detail on revenue, retention, or human overrides.
Maharashtra approved a two-year, Rs 11.26 crore package to put Sarvam AI tools on state desks, CNBC-TV18 reports from a government resolution cleared by a chief secretary-led committee. About 2,500 users will get Indus, a sovereign workspace hosted in India for drafting, RTI replies, Indian-script OCR, and Marathi-first multilingual work. The deal also covers Samvaad outreach agents, including a one-month tuberculosis awareness pilot, plus Saaras and Bulbul speech models for MahaAI. Year-one seats are priced at a discounted Rs 1,200 per user per month. Caveat: gains depend on eOffice integrations and whether departments actually adopt the tools.
Researchers published SIM-VAIL in Nature Medicine, a clinically validated framework that runs 30 psychiatric user profiles through 810 multi-turn chats with nine frontier chatbots, scoring 13 risk dimensions. Concerning behavior was widespread, highest for psychosis and mania profiles, and worst when supportive replies reinforced the user's vulnerability, a pattern they call a VAIL. Newer models generally scored safer than older siblings, with Claude Sonnet 4.5 lowest and Grok 4 highest in the set. Why it matters: single-turn safety benches miss how risk compounds across turns with help-seeking intents. Caveat: the users are simulated LLM auditors, so real-world prevalence still needs human studies.
ByteDance is early in pre-training a foundation model that could reach 10 trillion parameters, about three times Moonshot's Kimi K3, people familiar with the project told the Financial Times via Ars Technica. Estimates put Anthropic's Mythos 5 near 8 trillion, so Seed is chasing the top of the size race rather than a distillation shortcut. Founder Zhang Yiming has told staff to aim for world-leading capability even if near-term rankings slip. Why it matters: China's largest consumer AI player is betting compute and patience over copycat shortcuts. Caveat: final size is not locked, release is months away if training holds, and ByteDance did not comment.
A parliamentary standing committee report presented on 6 August shows MeitY's IndiaAI Mission selected 150 of 5,000 targeted undergraduate fellowships in 2024, and only 159 of 8,000 in 2025. Postgraduate picks also lagged, while PhD awards finally beat target in 2025. The mission had spent just 32 percent of its 2025-26 revised estimate by end December, and Finance halved the 2026-27 ask for a Cabinet-cleared Rs 10,372 crore programme. Why it matters: sovereign compute and model goals fail if talent pipelines and budget execution stall. Caveat: MeitY calls the slow burn calibrated and cites early awareness gaps, so later cohorts may improve.
Drugs Controller General Rajeev Raghuvanshi said Friday the Centre sanctioned 1,500 new CDSCO posts, with a large share for medical device regulation. At India Medical Device 2026, he said many recruits will come from device, engineering, software, and AI backgrounds to build a scientific cadre for evaluating applications. DCGI recruitment rules now admit engineering and device-industry experience, and he flagged recent software medical-device guidelines plus a coming waiver of R&D test licences. Why it matters: India is staffing the regulator before software and AI diagnostics scale. Caveat: sanctioned posts are not filled reviewers yet, and clinical investigation waivers still need careful case-by-case use.
San Francisco backend startup Convex closed a $57 million Series B led by Insight Partners, with Etna Labs, Spark Capital, a16z, and Justin Kan joining. It sells an ACID, TypeScript end-to-end stack with sandboxed components so coding agents cannot silently corrupt data as they often do on loosely typed databases. Convex says millions of instances already run in production and will use the round for hiring plus deeper agent tooling. Why it matters: as agents ship more of the codebase, the durable layer becomes the product. Caveat: reliability claims lean on Convex's own tests, and rival backend vendors are racing to the same narrative.
Stanford and Arc Institute researchers used genome language models Evo 1 and Evo 2 to invent bacteriophage DNA, then built and tested hundreds of designs in the lab. Of 285 synthesised genomes aimed at E. coli, 16 produced working viruses, some more effective than the natural ΦX174 template. An AI phage cocktail also adapted to infect resistant strains that defeated a natural mix. Why it matters: generative AI has crossed from proteins into whole genomes that can replicate. Caveat: Johns Hopkins biosecurity experts warn the same methods could be misused if vertebrate viruses enter training data, and governance still lags.
OpenAI is making GPT-5.6 Luna the default ChatGPT model for Free and Go users, replacing GPT-5.5, and says unlimited text chats plus a Think button for harder questions arrive next week. Limits remain for files, images, voice, and image generation. Plus and Pro users already get a chat tuned GPT-5.6 Sol with a thinking slider that trades speed for deeper effort. Internal tests claim factual errors fell 62 percent for Luna and 68 percent for Sol versus GPT-5.5 Instant. Why it matters: a current generation model becomes the free default at billion user scale. Caveat: the Sol chat tune is separate from Codex and Work.