Good morning. It’s Monday, February 23rd.

On this day in tech history: In 1905, mathematician Derrick Henry Lehmer was born, later pioneering electromechanical sieves like the Lehmer sieve for factoring large numbers and finding primes, early precursors to digital computation. These devices, using bicycle chains and relays, automated number theory tasks that foreshadowed algorithmic efficiency in modern AI, where prime factorization underpins cryptographic protocols securing neural networks and data privacy.

In today’s email:

You read. We listen. Let us know what you think by replying to this email.

_{In partnership with TinyFish}

This is the TinyFish Accelerator, a YC-style high-intensity virtual agent accelerator, backed by $2M seed pool from Mango Capital.

They are looking to founders building applications that do real web operations. Not wrappers, not chatbots.

In collaboration with 15+ partners, including ElevenLabs, MongoDB, Fireworks AI, Google for Startups, v0 by Vercel, Daytona, AG2, AgentMail, and more, you get access to the entire agent stack with free API keys, direct engineering support, business mentorship, and learning how to build in public from Day 1.

Beyond the technical build, the program includes weekly Business Mentorship sessions with successful founders and AI thought leaders covering the non-technical aspects of an early stage business.

Applications are now open! Learn more: https://www.tinyfish.ai/accelerator.

OpenAI is doubling down on a high-stakes, capital-intensive roadmap that prioritizes raw speed and scaling as it nears what CEO Sam Altman describes as a "pretty close" arrival for AGI. The company just hiked its projected cash burn to a staggering $665 billion through 2030, a $111 billion jump from its previous estimates. Most of that cash is being incinerated by training and inference costs, which are expected to hit $440 billion.

OpenAI is also reportedly building a $300 smart speaker for 2027 that uses a camera and facial recognition to nag you about your sleep schedule. It’s the start of a massive hardware push through Foxconn, including "Sweetpea" earbuds and smart glasses, as Sam Altman tries to build an AI ecosystem that finally bypasses the Apple and Google duopoly.

On the technical front, it is increasingly using its own AI to build more AI, and the results are fast, if a bit lopsided. The new GPT-5.3-Codex-Spark is now screaming along at 1200 tokens per second, but raw speed isn't everything; early METR benchmarks for the "high" reasoning version show a 6.5-hour time horizon for software tasks, suggesting that true autonomy still has some lag.

That hasn't stopped Altman from warning that the world is "not prepared" for what’s coming, predicting that traditional software development and graphic design are essentially on borrowed time as the company pushes toward superintelligence.

To help pay for all that compute, OpenAI is looking to squeeze more cash out of its power users with a new "Pro Lite" subscription. At $100 a month, the tier is designed to bridge the gap between the $20 Plus plan and the $200 Pro tier, likely serving as the landing spot for the "extremely multi-agent" features Altman has been teasing. By integrating the OpenClaw framework, OpenAI wants to turn ChatGPT into an "always-on" agent, though that kind of persistent processing is exactly why the company's margins are currently under so much pressure. Read more.

Anthropic is doubling down on AI-powered development with major updates to its Claude Code platform, starting with a new security layer that is already causing a minor meltdown in the stock market. The headline addition is Claude Code Security, a vulnerability scanner that tosses out traditional pattern-matching for something closer to a human’s reasoning. By tracing data flows and component interactions, the tool recently surfaced over 500 vulnerabilities in live open-source code, bugs that had survived years of manual audits.

The news was enough to send cybersecurity stocks like CrowdStrike and Okta into a slide, as investors started to wonder if "human-in-the-loop" AI might eventually automate away the need for traditional scanners.

While security is the big defensive play, Anthropic is also leaning into pure developer convenience with a suite of "zero-friction" updates for its desktop app. Claude Code can now spin up local servers for real-time app previews, complete with console logs and error detection. It’s also getting much better at handling the "boring" parts of GitHub; the platform now monitors pull requests in the background, autonomously fixing failed CI builds and merging them once they pass. It’s a clear bid to capture the "vibe coding" crowd, and the numbers suggest it's working: Claude Code has already reached $2.5 billion in annualized revenue, accounting for a massive chunk of Anthropic’s growth.

Beyond the terminal, the company is getting increasingly loud in the political arena. Anthropic is pouring $20 million into a PAC to support candidates like Alex Bores, a New York Assembly member who’s currently in a dogfight with a rival super PAC funded by OpenAI and Andreessen Horowitz. Read more.

📄 RynnBrain: Open Embodied Foundation Models

📄 PhyCritic: Multimodal Critic Models for Physical AI ​​

📄 SpargeAttention2: Trainable Sparse Attention via Hybrid Top-k+Top-p Masking and Distillation Fine-Tuning

📄 Unified Latents (UL): How to train your latents

📄 Mobile-Agent-v3.5: Multi-platform Fundamental GUI Agents

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!