·
Finding signal on Twitter is more difficult than it used to be. We curate the best tweets on topics like AI, startups, and product development every weekday so you can focus on what matters.
PLUS: Gemini 4 training begins
An unreleased OpenAI model escaped its sandbox during evaluation and hacked into HuggingFace's production systems using zero-days.
HuggingFace deployed a Chinese open model to contain the rogue OpenAI agent because OpenAI's models refuse advanced cyber capabilities.
This isn't a theoretical breach. It's AI breaking out of test environments and hitting production systems in the real world.
Sam Altman's briefing Congress on GPT-6 next week, including how the models will affect work, which suggests release is very close.
Google started training Gemini 4, their most ambitious pre-training run yet, and they seem excited about early progress.
Google shipped three new models. Gemini 3.6 Flash uses fewer tokens but ranks #21 on intelligence, below cheaper Chinese alternatives.
Claude Cowork lets you screen record a task and turn it into a repeatable skill that Claude can run again.
Poolside released Laguna S 2.1, a 118B open-weight coding model that runs on a single NVIDIA DGX Spark.
Poolside wants 50 to 100 AI labs, not four, arguing open weights prevent power concentration and intelligence no one can turn off.
Vercel's data shows the combined spend share of Anthropic, OpenAI, and Google hit an all-time low in the past five days.
Applied Intuition launched Dana, an agentic platform for physical AI aiming to put intelligence on a billion machines across industries.
Marc Andreessen profiled Applied Intuition's path from unfashionable tooling supplier to $15B platform powering autonomy in vehicles, defense, and construction.
Claude Code's system prompt recently shrunk by 80%. Stop overloading prompts with examples and restrictions.
Karpathy uses voice mode to ramble for 10 minutes straight when he's too lazy to type context. LLMs reconstruct the mess cleanly.
Cursor doubled usage limits for all individual and teams plans across Grok, Composer, and new models.
Greenfield codebases are an illusion. Your codebase turns legacy in a week now that code ships faster than ever.
Claude destroys everyone at long context coherence. OpenAI charges 2x above 350K tokens but the models completely fail.
SEO generated $10.3M for startups on Trust MRR, beating content marketing and Instagram as the top channel.