Inherent, founded by DeepMind alumni, says its AI ’teammate’ just outperformed Anthropic and OpenAI at replicating research — TechCrunch AI London-based Inherent, backed by a reported $50M seed led by Index Ventures, released Faraday — a compact 27B-parameter “AI Scientist” built atop Qwen 3.6 and trained with long-horizon reinforcement learning. On Inherent’s new Replica benchmark (310 tasks drawn from 100 ML and AI-for-science papers spanning NLP, materials science, and weather forecasting), Faraday outperforms Claude Opus 4.8 and GPT-5.5 in every category, with the most pronounced advantage in structural biology and materials science.
🔬 Frontier Models & Research
Inherent, founded by DeepMind alumni, says its AI ’teammate’ just outperformed Anthropic and OpenAI at replicating research — TechCrunch AI
Faraday is trained to replicate scientific figures from papers without access to the original plots — a task that serves as a stepping stone toward open-ended AI-driven discovery. The Replica benchmark covers 310 tasks from 100 real ML and AI-for-science papers; Faraday beats both frontier models in every category, with the largest gaps in meta-learning, structural biology, and materials science. Inherent’s $50M seed round also included Radical Ventures, and this is the company’s first major public release since emerging from stealth.
OpenAI cuts flagship model pricing by over 20% — AI News — Sol now costs $4 per million input tokens and $20 per million output tokens, with promotional pricing locked in through at least November 2026.
DeepSeek V4-Pro graduates to full production — AI News — DeepSeek moved V4-Pro out of limited preview this week, introducing peak and off-peak billing tiers for the first time.
Alibaba publishes open weights for its latest model — AI News — The openly downloadable version is text-only; the full hosted model additionally handles images and video.
Z.ai’s GLM-5.3 open coding model finds over 1,000 critical bugs in widely-used software — AI News — The open-source model’s large-scale bug-hunting run highlights the growing role of specialized AI coding agents in automated vulnerability discovery.
⚖️ Policy, Legal & Geopolitics
Is it legal to train AI models on copyrighted books? It’s complicated — TechCrunch AI
Courts are reaching contradictory conclusions: two district judges found AI training constitutes fair use on their records; a Third Circuit judge found against it. Anthropic’s $1.5B Bartz settlement — which received final court approval last month, covering 482,000+ works — established that training on books can be fair use while storing pirated copies cannot. The piece argues the most potent future argument against AI companies may be market harm: if a model trained on your books can flood the market with similar work and undercut your livelihood, that is precisely the damage fair use analysis is designed to prevent.
Washington considers asking countries to choose between US and Chinese AI ecosystems — AI News — The US-China AI rivalry is taking a new geopolitical turn, with Washington reportedly weighing pressure on dozens of allies to align with either American or Chinese technology stacks.
🏢 Enterprise & Venture
Harvard’s $699 startup bootcamp offers AI avatars of its instructors — TechCrunch AI — The HBS Foundry program deploys AI avatar versions of instructors to give real-time feedback during practice pitch sessions and simulated board meetings.
Flock CEO calls for ‘compromise’ as surveillance company faces growing backlash — TechCrunch AI — Flock Safety’s automated license plate reader network is drawing sustained public scrutiny over the potential for its surveillance infrastructure to be misused.
Will the DOJ’s investigation into a16z spook other VCs? — TechCrunch Venture — The Justice Department is examining how startup board seats are structured at the major AI-focused venture firm, prompting broader questions about VC governance across the industry.
Generated by claude-sonnet-4-6 · 2026-08-23T10:00:00Z