Meta Ships Muse Wonder, Its First Frontier Model -- $4/$20 on the API, the Bigger Version Kept Inside Muse Code, and 71 of 92 Cyber Flags Taken the Unintended Way
Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12, positioned for long, multistep agent work. Meta's own model card shows a split scorecard against Claude Opus 5.5 and GPT-6 Astra, and a safety section that does not rule out the high-risk chem-bio threshold without mitigations.
HCLTech's AI Revenue Jumps 78% to $219 Million a Quarter -- and Its CEO Says AI-Led Price Deflation Is Only in Its 'Early Phases'
HCLTech reported September-quarter results on October 12 that put numbers on both sides of AI for Indian IT services: a fast-growing new AI business, and a core services book where productivity gains are cutting what clients pay. The company raised the bottom of its FY27 guidance anyway.
'Adception': A Google Ad That Pointed at Bing Sent 'claude mac' Searchers to a Fake Installer Whose Copy Button Swapped the Command
Push Security reported on October 9 a malvertising chain that got past Google's ad review by listing bing.com as the destination, then used Bing's click redirect and a hacked retailer's page to reach a polished fake Claude download site. It is the latest in a run of attacks that target developers installing coding agents.
Research Radar: SoftwareGym Builds 182 Fake Desktop Apps to Train Computer-Use Agents; ProgramGym Turns 1,648 Repos Into Rebuild-From-Scratch Environments
Two papers near the top of Hugging Face's daily list on October 12 attack the same bottleneck from different sides: agents need many verifiable environments to train in, and building them by hand does not scale. Both generate the environments automatically and check them before use.
Framework Watch: LangGraph 1.2.15 Stops Saving Checkpoints Past a Failed Write and Stops ToolNode Swallowing Human Answers; LangChain Moves 'Mid-Conversation' Support Into Model Profiles
LangGraph 1.2.15, released October 12, is almost entirely persistence and interrupt fixes, several of which silently lost state or skipped human checkpoints. The same day's LangChain releases replace model-name matching with declared capabilities, and let summarization run on a different model.
Google's Three-Month Trial With 133 Patent Lawyers: AI Made Everyone's Drafts Better, but Only Senior Lawyers Got Better Without It
A randomized trial led by MIT economist David Autor at Google Research, published October 7 with an NBER working paper, finds AI drafting help raised quality for junior and senior lawyers alike. On an unassisted task afterwards, only seniors showed lasting gains.
Trend Watch: Karpathy Says About 5,000 Insiders See 'the Curve Going Vertical' -- Two Days After a Developer Asked 'Why Are Coding Agents So Dumb?'
Andrej Karpathy's October 11 post sorting the world into a funnel of LLM experience drew more than 5 million views. It landed beside Michael Lynch's widely discussed complaint that the models improved while the agent software wrapped around them did not. Both can be right, and the gap between them is the story.