-
release · Google DeepMind · 2건
EmbeddingGemma 2: an open, lightweight multimodal embedding model
Google DeepMind introduced EmbeddingGemma 2, an open, lightweight multimodal embedding model. Simon Willison praised its Apache 2.0 license, arguing that closed, hosted-only models make little sense for embeddings.
If you store thousands or millions of vectors for search or RAG, Apache 2.0 weights remove the risk of paying to re-embed everything when a vendor retires a proprietary model.
-
product · The Verge AI
Google is about to remove free access to Gemini Flash and Pro
Starting October 9th, Google will limit free-plan Gemini users to the Flash Lite model, with the standard Flash requiring the $4.99/month Google AI Plus subscription. AI Plus is also losing Gemini Pro, leaving Pro and the "Deep Think" reasoning option to Google AI Pro or Ultra subscribers.
Developers who lean on Gemini's free tier for prototypes or personal tooling should expect Flash Lite-level output after October 9 and weigh the cost of paying for a subscription versus moving to another model.
-
research · Hacker News
Opus 5.5 agents discover two room-temperature magnetic semiconductor candidates
A Hacker News post reports that Opus 5.5 agents discovered two candidate materials for room-temperature magnetic semiconductors. The materials are described as candidates, not confirmed discoveries.
Developers who give agents long-running science tasks should separate candidate generation from experimental validation and budget for the validation cost.
-
research · Hacker News
Dust: Pretraining Transformers Without Backpropagation
A post titled "Dust: Pretraining Transformers Without Backpropagation" appeared on Hacker News. Its title indicates an approach to pretraining Transformers that does not use backpropagation.
If pretraining without backpropagation holds up, it could change activation-memory needs and training pipeline design, so anyone building training infrastructure should check the method's scale and benchmark results first.
-
product · Anthropic
Expanding the Cyber Verification Program
Anthropic announced it is expanding its Cyber Verification Program. The headline alone does not spell out the scope or terms of the expansion.
Developers using Claude for security work should check Anthropic's announcement for changes to eligibility and coverage before settling on tooling.
-
product · Simon Willison
Quoting Felix Rieseberg
Anthropic's Felix Rieseberg says the old Cowork ran inference in the cloud but executed tool calls in a VM shipped to the user's computer, which cost disk, battery and performance and stopped work when a laptop was closed. The new version runs both inference and the VM in the cloud, gives each session its own sandbox, and has the desktop app handle file access tool calls when the VM needs something on the device.
For anyone designing agent runtimes, it shows the trade-off between local and cloud sandboxes: battery cost, whether work survives a closed laptop, and how device files get accessed.