★ Top Story
MIT Technology ReviewAnthropic unveiled Claude Science at an event for pharmaceutical executives, biotech founders, and researchers — positioning it as a domain-specific AI workbench for science in the same way Claude Code serves software engineering. The product can autonomously execute complex research workflows from high-level instructions, integrating databases, pipelines, and computational tools into a single environment. NVIDIA's BioNeMo Agent Toolkit is among the first integrations, bringing GPU-accelerated life sciences capabilities directly into the platform.
Read the original →
TechCrunchAfter weeks of negotiations with the Trump administration — which had briefly spooked the White House into mandating safety reviews — Anthropic's most advanced Fable 5 and Mythos models are now cleared for global deployment. Access is being restored across Claude's own platforms as well as AWS, Google Cloud, and Microsoft Azure. The episode highlights the continued uncertainty surrounding U.S. AI export and safety policy, with industry observers noting the administration's approach remains inconsistent.
Read more →
TechCrunchAnthropic released Claude Sonnet 5, a mid-tier model designed to deliver strong agentic performance at a lower cost than flagship offerings like Opus, GPT-5.5, and Gemini Pro. The model features improved safety guardrails and is aimed at developers building autonomous agents who need to balance capability with inference economics. The release reinforces Anthropic's strategy of maintaining a tiered model lineup to capture both enterprise and cost-sensitive developer markets.
Read more →
TechCrunchMeta is developing plans to monetize its massive AI infrastructure by offering compute access and model APIs to external customers, putting it in direct competition with AWS, Google Cloud, and Microsoft Azure. The move mirrors a similar strategy reportedly being explored by SpaceX, reflecting a broader trend of AI-heavy companies treating surplus compute as a revenue stream. If executed, it would mark Meta's most significant push into enterprise cloud services to date.
Read more →
NVIDIA BlogNVIDIA is inviting capital and infrastructure partners to co-build large-scale, multi-tenant AI factories designed for continuous production inference — the phase of AI deployment that generates tokens at scale rather than training new models. The program reflects a strategic shift in the AI compute market, where utilization economics and time-to-deployment are becoming as important as raw chip performance. NVIDIA is positioning itself as the full-stack provider for this next wave of AI infrastructure buildout.
Read more →
TechCrunchCloudflare is drawing a hard line between web crawlers used for search indexing and those used for AI training or agent retrieval, requiring AI companies to distinguish between the two by September 15 or risk being blocked by default across a large swath of publisher sites. The policy is designed to give content creators and publishers leverage to negotiate compensation from AI companies that scrape their work. It represents one of the most concrete enforcement mechanisms yet in the ongoing battle over AI training data rights.
Read more →
TechCrunchEtched, a startup building inference-optimized AI chips designed to challenge NVIDIA's dominance, has reached a $5 billion valuation and announced $1 billion in contracted sales for its inference systems. The milestone signals growing enterprise appetite for alternative AI silicon, particularly for cost-efficient inference workloads where purpose-built chips can outperform general-purpose GPUs. Etched's traction underscores the intensifying competition in the AI chip market beyond the training segment.
Read more →
TechCrunchVenice AI, which offers an AI platform built around user privacy and local inference, has closed a $65 million Series A at a $1 billion valuation — and is already generating over $70 million in annualized revenue. CEO Erik Voorhees attributed the growth to rising consumer distrust of data-hungry AI platforms and demand for alternatives that don't train on user inputs. The company's profitability at this stage is unusual in the AI startup landscape and strengthens its position heading into a crowded market.
Read more →
MIT Technology ReviewA new MIT Technology Review investigation highlights a persistent and underappreciated flaw in large language models: they converge on statistically dominant responses, producing predictable outputs that undermine their usefulness for tasks requiring genuine diversity of thought. A startup profiled in the piece is developing techniques to inject controlled stochasticity and perspective diversity into LLM outputs without sacrificing coherence. The problem has significant implications for AI use in research, creative work, and decision-making where echo-chamber reasoning can compound errors.
Read more →
TechCrunchGoogle has expanded Gemini Spark — its always-on agentic desktop assistant — to macOS, adding real-time activity tracking and support for a broader range of third-party applications. The Mac launch positions Gemini Spark as a direct competitor to similar ambient AI assistants from Microsoft and Apple, targeting power users who want continuous AI assistance across their workflow. The expansion is part of a broader push detailed in Google's June 2026 AI update roundup.
Read more →
TechCrunchSpaceX reportedly demonstrated a "handset-like" AI device to investors ahead of its anticipated IPO, fueling speculation that the company is eyeing a move into consumer wireless hardware. The prototype's existence, combined with SpaceX's Starlink connectivity infrastructure and its reported plans to monetize excess compute, suggests a broader ambition to build a vertically integrated AI communications ecosystem. No product announcement has been made, but the investor demo signals the concept is further along than previously known.
Read more →
Ars TechnicaSecurity researchers have demonstrated a novel attack against AI-powered browsers in which feeding an LLM false premises — such as asserting that 2+2=5 — is sufficient to destabilize its internal reasoning context and cause it to follow otherwise forbidden instructions. The attack raises serious concerns about the safety architecture of agentic browsers that are granted access to sensitive user data and web actions. Researchers argue the findings are one more reason why deploying LLMs as autonomous browser agents introduces systemic risks that current guardrail designs cannot reliably contain.
Read more →
Hugging FaceHugging Face and Cerebras have partnered to deploy Google's Gemma 4 model for real-time voice AI applications, leveraging Cerebras' wafer-scale inference hardware to achieve the low latency required for natural spoken interaction. The collaboration demonstrates a growing ecosystem around open-weight models being adapted for specialized, latency-sensitive deployment scenarios beyond text. It also highlights Cerebras' continued push to position its hardware as the go-to platform for real-time generative AI workloads.
Read more →
TechCrunchSerial entrepreneur Bhavin Turakhia is personally funding Neo, his fifth venture, with $30 million to build an AI-native productivity suite designed to displace Microsoft Office and Google Workspace. Turakhia's previous ventures include Radix and Flock, and he is betting that the transition to AI-first workflows creates a rare opening to unseat entrenched incumbents in enterprise productivity. The self-funded approach gives him runway to build without immediate investor pressure, though competing with Microsoft's deeply integrated Copilot ecosystem will be a formidable challenge.
Read more →
TechCrunchUK-based autonomous driving AI company Wayve has launched an $85 million employee tender offer, valuing the company at $8.5 billion and reflecting continued investor confidence in its embodied AI approach to self-driving. The tender is part of a broader trend among well-funded AI startups using secondary liquidity events to attract and retain top engineering talent in a competitive hiring market. Wayve's software-defined, data-driven approach to autonomy has distinguished it from hardware-heavy competitors.
Read more →
TechCrunchEquiLibre Technologies, founded by three former DeepMind researchers who previously built a world-class poker-playing AI, has surpassed a $500 million valuation by applying game-theoretic AI to quantitative finance. The Prague-based lab is now generating returns for hedge fund clients by using multi-agent reinforcement learning techniques originally developed for adversarial games. The company's trajectory illustrates how foundational AI research in strategic reasoning is finding lucrative commercial applications in financial markets.
Read more →
The VergeGoogle is rolling out a new NotebookLM feature that generates 60-second vertical AI video clips summarizing uploaded research sources, borrowing the short-form format popularized by TikTok for knowledge consumption. The feature is initially available to Google AI Ultra and Pro subscribers and represents a significant expansion of NotebookLM beyond its audio podcast summaries. It signals Google's intent to make AI-assisted research synthesis accessible across multiple media formats.
Read more →
The VergeNetflix's upcoming Willy Wonka reality competition, "Wonka's The Golden Ticket," will feature an AI-generated voiceover modeled on the late Gene Wilder, raising immediate ethical questions about posthumous likeness use in entertainment. The show, premiering September 23rd, follows Netflix's Squid Game reality format of turning fictional scenarios into real competitions. The use of a synthesized celebrity voice without the subject's living consent is likely to intensify ongoing industry and regulatory debates around AI voice cloning rights.
Read more →No stories from this source today.