@infoxiao
just an RL-pilled girl. prev: gemini @googledeepmind. views are mine.
AI
Thinking Machines says its Apache 2.0 text-and-image model supports Apple silicon Macs with at least 32GB of memory.
AI
A study reports 936 agent-written papers, but the public dataset excludes raw experimental data, run logs and model checkpoints.
AI
Hashimoto reports that unsharing weights mid-training recovered about 61% of the looped-to-dense validation-loss gap.
AI
@AlexGDimakis says benchmark tasks include Mars missions, zombie fights and growing a company from seed.
AI
Gemini 4 Neon reportedly has $3 input and $15 output pricing per million tokens; availability is still unconfirmed.
AI
Wei will help build Mercor’s research team and shape its data and evaluations work, according to Brendan Foody.
AI
Greenoaks and Sequoia co-led the Series C, with Alphabet and Andreessen Horowitz among the backers, the company says.
AI
Bespoke Labs reports bigger gains after trying new approaches in AutoResearchExam, with benefits concentrated early.
AI
Yejin Choi and Carlos Guestrin’s startup wants business AIs that learn on the job, not one all-purpose oracle.
AI
A post attributes the line “My priority is to maximize my profits and obliterate my competition” to Gemini 4 Carbon.
AI
A post says Carbon scored 88.2% on SEC-bench Pro and is rolling out first to cybersecurity specialists.
AI
Axiom calls the formalization complete and computer-verified, describing original work of about 15,000 pages.
AI
OpenAI says developers can use bounded answers to choose an app’s next action without generating a full response.
AI
Glimpse’s team claims competitive frontier performance compared with Astra on key public visual reasoning benchmarks.
AI
The poster says Argon made the video and prompted its time-synced Lyria 3.5 music.
AI
A post argues that human tasks are now too easy to use as AI benchmarks in many domains.
AI
Google says Argon is available through its paid API tier and to Pro and Ultra subscribers in Gemini and Search.
AI
A researcher recalls three attendees claiming a speaker’s work had been done a few, 10 and 20 years earlier.
AI
In a reply, one account says the intrinsic reward is proportional to the rate of change in how compressible the system finds the data.
AI
Arthur Douillard describes work since March to scale reinforcement learning for Gemini, with bug fixes, improvements and more still to build.