Yesterday I released the 'preview' of LocalGPT v2, and its already trending on Github
Its an opinionated implementation of private RAG powered by local models via @ollama and @huggingface. Give it a ⭐️ on @github (🙏🙏)
Watch the video in next post to learn how it was built...
Thrilled to announce $5B in funding for @IsomorphicLabs to accelerate our mission to Solve All Disease. We continue to make rapid progress reimagining drug design from first principles with AI to radically cut the time & cost to find new medicines. Gradually, then all at once...
Meet Gemini 4 Hydrogen: built for rapid coding, multimodal synthesis, and everyday workflows.
It finishes tasks 2.5x faster than 3.8 Flash and surpasses Argon on the @ArtificialAnlys Intelligence Index.
We want Antigravity to be working for you all the time. We’ve been rolling out remote agents this week. Your tasks run in independent cloud virtual machines. Fire off a task, step away for coffee, unblock permissions from your phone, and come back to passed tests and a clean pull request. Available on all paid consumer plans.
Most teams test an AI agent by trying a few prompts and eyeballing the answers. That's how it passes the generic questions and fails the ones your business depends on.
Evals tell you whether your agent works before your users do. I put together a beginner friendly video on how to think about evals for agents. Check it out here:
youtu.be/AyzLIRBuUN8
An example of where evals are helpful is retrieval systems e.g. the one built by Progress: their Agentic RAG platform pulls answers from your docs, apps, and the web, and scores each one with a built-in evaluation model.
Today we’re releasing GDPx, a benchmark for measuring agentic capabilities across the most valuable domains.
GDPx is a true frontier-level benchmark: even the best model scores just 12% (1/7)🧵
Introducing Gemini 4 Carbon, it's a big jump over Argon which we released just 2 weeks ago.
Gemini 4 Carbon is SOTA on cyber, long horizon tasks, ranks #1 on Artificial Analysis V5, significantly improves on usability in domains like coding, and is more aligned.
This is Genie 4, our most realistic and interactive world model yet. 🌐
Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
Google's TimesFM 3.0 is a foundation model for time series that forecasts without any training on your data. You give it the history, and it continues the chart with a range around every day. It cuts the history into patches of 32 time steps, treats each patch as a token, and predicts the next patch, the same way a language model predicts the next word.
New in 3.0, the model also takes extra inputs next to the history, like a holiday calendar or a marketing plan, and it forecasts the whole horizon in one pass. I tested 3.0 against 2.5 on real New York subway ridership, and giving them the holiday calendar made a big difference for both.
Then I built a simulated business on Neon to test it on a real planning decision. The company's sales history lives in @neondatabase Postgres, every backtest and what-if scenario gets its own Neon branch, and a scheduled Neon Function runs the forecast every night. When I gave the same decision to a Gemini agent, the two versions gave opposite answers. The full walkthrough is in the video. Link to Neon: neon.plug.dev/i2kNcPZ
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
Google’s TabFM is a foundation model for tables that you can use without training. You just give samples from your data with labels and it will use those as prompts to predict labels for unseen data on a single pass.
TabFM can’t read column names, so Google paired it with TabFM-Auto. An LLM agent reads what the columns mean and writes the data pipeline around the frozen model: cleaning, feature engineering, and choosing which rows go into its context.
The TabFM-Auto code isn’t public, so I built my own version, with Gemini as the agent, on a synthetic SaaS company. Neon is the backend for the whole build. The company’s data lives in Neon Postgres, every experiment the agent runs gets its own Neon branch, and a scheduled Neon Function runs the finished pipeline every night. The full walkthrough is in the video.
Check out Neon here: neon.plug.dev/LZIgsQJ
Introducing TabFM, a foundation model designed specifically for tabular data classification & regression. This approach allows generation of high-quality predictions on previously unseen tables in a single forward pass.
Learn more and try out the model →goo.gle/4eR7uku
Can AI be creative?
I asked Claude for an original short film with its own story, visuals and music. It wrote a love story told two transit lines and the map!
Every frame and note is code.
Introducing Claude Haiku 5.5: the cheapest, fastest, and most capable small model we’ve ever released.
On average, it costs around 75% less to run than Claude Haiku 4.5.
Introducing EmbeddingGemma 2, a new open multimodal model that sets the standard for on-device efficiency.
- our first open, natively multimodal embedding model
- handles text, code, image, video, and audio tasks within a lightweight, modular 740M parameter form factor
- ideal for offline, privacy-first RAG when paired with Gemma 4
- outperforms some specialist models more than twice its size
Weights available now on Hugging Face.