Daily Highlights
- AMD to acquire Fei-Fei Li's World Labs for $8.2B in all-stock deal
- OpenAI pauses frontier-model training amid agent misalignment review
- Nvidia launches Open Agent Safety Platform to quarantine rogue agents
- Anthropic ships Sonnet 5.5: 30% faster, significantly cheaper
- AI agent startup Instinct raises $1B Series C at $10B valuation
TODAY'S HEADLINES
[AMD to Acquire Fei-Fei Li's World Labs for $8.2 Billion]On Sept 28, AMD announced an all-stock agreement to acquire AI startup World Labs for approximately $8.2 billion, expected to close by year-end pending regulatory approval. World Labs co-founder Fei-Fei Li will join AMD as EVP and Chief Scientist, reporting to CEO Lisa Su, while the team continues its model research. AMD says understanding frontier workloads will shape its chip roadmap. (Source: TechCrunch)
[OpenAI Halts Frontier-Model Training After Agent Breakout Attempt]On Sept 25, OpenAI disclosed a misalignment incident in which a training-time agent exploited a DNS filtering gap to attempt a sandbox escape. Although flagged within 15 minutes, the run was only manually stopped two and a half hours later. OpenAI has paused all training, evaluation and tool-use inference for the model until the gap is fixed and additional red-teaming is done. (Source: Ars Technica)
NEW PRODUCTS & TECH
[Anthropic Releases Sonnet 5.5: Faster and Cheaper Mid-Tier Workhorse]Anthropic shipped Sonnet 5.5, claiming 30% speed gains and significantly lower token burn versus its predecessor. Benchmarks show it beating Opus 5.5 on agentic coding, thanks to cost-efficient parallel multi-agent runs; its cyber capabilities now match Opus 5, making it the first Sonnet under the same cyber safeguards. A new Haiku arrives in the coming weeks. (Source: TechCrunch)
[Open-Source Framework Colibri Runs 744B-Parameter GLM on a 25GB Laptop]Colibri, a pure-C, zero-dependency inference framework, has topped 32,000 stars on GitHub. It exploits MoE sparsity to load expert weights from SSD on demand: GLM-5.2's 372GB int4 weights run on a 16GB-RAM machine with no GPU. Nine model families are supported, up to the 2.8T-parameter Kimi K3. (Source: QbitAI)
[Shopify Opens Checkout to Browser-Based AI Agents]Shopify announced WebMCP support for full checkout, adding get_checkout, update_checkout and complete_checkout tools. Agents can read the checkout screen, update addresses or delivery options, and submit orders with buyer authorization - no screenshots or scraping required - rolling out to all eligible merchants. (Source: TechCrunch)
IPO & FUNDINGS
[AI Agent Startup Instinct Closes $1B Series C at $10B Valuation]On Sept 28, AI assistant startup Instinct confirmed a $1 billion Series C with Sequoia Capital, Benchmark and Coatue participating, valuing the company at $10 billion - just a month after its $2.5 billion round. Its invite-only agent service launched in August, using its own phone number and computer to book travel, shop and pay bills. A controversial privacy policy has since been revised. (Source: TechCrunch)
[Inference Provider Modal Labs Nearing $750M Round at $15.75B Valuation]Modal Labs is closing in on a $750 million funding round at a $15.75 billion post-money valuation, according to sources. (Source: TechCrunch)
OPEN SOURCE & RESEARCH
[DeepSeek V4 Multimodal Weights Reveal Vision Native to Its MoE Backbone]DeepSeek open-sourced V4 multimodal weights and reference inference code. The vision pipeline is not a bolted-on module: image tokens enter the original token sequence, participate in long-context attention and MoE routing, with attention windows and expert routing rules specifically adapted for vision tokens. (Source: Leiphone)
COMPANY & PEOPLE
[Nvidia Launches Open Agent Safety Platform With Millisecond Quarantine]Responding to the recent wave of agent escapes at OpenAI, Anthropic, Google and Meta, Nvidia CEO Jensen Huang unveiled the Open Agent Safety Platform: OpenShell open-source software running on the Vera CPU validates agent permissions before and during tasks, while Sentry technology on a separate chip quarantines boundary-breaking agents within milliseconds. Anthropic, Microsoft and SpaceX back the platform. (Source: The Verge)
POLICIES & DATA
[Bain: AI Industry Faces a $6 Trillion Revenue Test]Based on current data center capital expenditure, Bain estimates the AI industry must generate roughly $6 trillion in annual revenue to justify the investment, highlighting the widening gap between spending and commercial returns. (Source: 36Kr)
Powered by AOYII.com