X / Twitter
Anthropic Claude
Claude Code has officially introduced computer use in research preview for Pro and Max users on macOS. Claude can now open applications, click their UIs, and test what it builds directly from the command line. Specifically, it supports any application that can open on a Mac, including compiled SwiftUI apps, local Electron builds, and GUI tools without CLIs. In a single prompt, Claude can write code, compile it, launch an app, click through the interface, find bugs, fix them, and verify the fixes. Enable it by running /mcp in Claude Code. That same day, Claude Code’s Auto mode became available to Enterprise and API users; update and run `claude --enable-auto-mode` to enable it.
https://x.com/claudeai/status/2038663014098899416
https://x.com/claudeai/status/2038693742094246032
Anthropic Claude Code engineer Cat Wu
Claude Code now supports GitHub Enterprise Server across its product line, including web, iOS, Android, and Code Review.
https://x.com/_catwu/status/2038723796178526712
Vercel CEO Guillermo Rauch
Following Opus 4.5’s release, engineering at Vercel changed irreversibly, with agents now doing the vast majority of coding. Guillermo said that precisely because Vercel understands LLMs’ inherent flaws and overconfidence, it sent its team a clear message: vibe coding and mission-critical infrastructure cannot be mixed. The company is sharing its early internal guidelines for responsible agent use, centered on always prioritizing security, durability, and availability.
https://x.com/rauchg/status/2038759092442050651
Box CEO Aaron Levie
In a world where agents process information hundreds or even thousands of times faster than people, existing infrastructure and services will require fundamental rebuilding. Levie noted that even tools designed for machine speed usually still serve a particular end-user request. That premise is now breaking down, and assumptions across the technology stack need to be revisited.
https://x.com/levie/status/2038468564500537416
Replit CEO Amjad Masad
Replit announced a partnership with RevenueCat: users can now add monetization directly to mobile apps built on Replit.
https://x.com/amasad/status/2038722068418805875
Swyx
Redpoint published a list of SaaS businesses suitable for rebuilding from scratch with AI. Swyx inverted one of its figures, pointing out that “46% of enterprise CIOs are open to AI-native startups replacing incumbent vendors” represents a remarkable market opportunity. He said he would not have reached that conclusion before seeing the survey.
https://x.com/swyx/status/2038431061575979027
Builder Zara Zhang
Her open-source Codebase-to-course project has reached 2600 GitHub stars. She optimized it to substantially improve token efficiency and stability. Originally designed to help vibe coders learn computer science, it has also proved highly effective for developer onboarding.
https://x.com/zarazhangrui/status/2038685338865221862
Every CEO Dan Shipper
He sharply criticized a company’s statement about the serious privacy breach of “users’ private data being mistakenly sent to other users,” calling the wording overly euphemistic, insincere, and an extremely poor way to communicate.
https://x.com/danshipper/status/2038688359325835436
Podcasts
Latent Space — Mistral: Voxtral TTS, Forge, Leanstral, & Mistral 4 — w/ Pavan Kumar Reddy & Guillaume Lample
Key takeaway: Mistral released its first TTS speech synthesis model, Voxtral, using a proprietary flow-matching architecture to achieve industry-leading performance in low-latency streaming generation. It also detailed Forge, an enterprise model-customization platform, and Leanstral, a project focused on formal mathematical proofs.
Pavan Kumar Reddy leads audio research at Mistral and previously worked on post-training research for Google Gemini. Guillaume Lample is Mistral’s chief scientist and previously helped release LLaMA at Meta. This launch showcased Mistral’s latest progress in audio, enterprise customization, and foundational research.
Voxtral TTS’s technical approach
Voxtral is a speech synthesis model built on a Ministral backbone with 3B parameters and support for nine languages. Its central innovation replaces the mainstream “depth transformer + discrete tokens” approach with a proprietary flow-matching output head. Specifically, the model’s neural audio codec converts audio into latent tokens at 12.5 Hz, or 12.5 frames per second, each containing semantic and acoustic tokens. After the main transformer decoder produces latents, they do not undergo discrete sampling. Instead, the flow-matching head estimates a velocity field, denoising noisy latents into audio latents, which a vocoder decodes into actual audio. Compared with the K autoregressive prediction steps needed by a depth transformer, flow matching requires only 4-16 inference steps, greatly reducing latency, and naturally supports further compression to single-step inference. Pavan explained that the fundamental reason for choosing flow matching is speech’s high entropy: the same word can be spoken with countless tones, pauses, and stresses, so the distribution must be modeled rather than predicting an average, which corresponds to blurred, distorted speech. Guillaume added that this is a relatively novel use of flow matching in audio generation. The image community is ahead, while audio still has no settled optimal approach, leaving substantial room for research. Voxtral’s very low latency suits real-time streaming scenarios such as voice agents, which Mistral considers the most important application direction for TTS.
Forge: enterprise model customization
Guillaume detailed Mistral’s core value proposition for enterprise customers. Companies accumulate vast amounts of vertical-domain data, sometimes reaching trillions of tokens, unavailable on the public internet and therefore inaccessible for closed models to truly exploit. Injecting context alone is far less effective than fine-tuning on that data. Forge is essentially the same toolchain Mistral’s internal science team uses for pretraining, SFT, and RLHF, now highly mature after two years of refinement. He offered examples: a pretraining mix containing 50% of a particular Asian language for one customer, giving the model a much stronger grasp of its dialects and idioms than general models; a voice interaction model that runs offline in a customer’s vehicle systems because connectivity is unreliable; and rebuilding customers’ applications that were prototyped with closed models but too costly to operate, using fine-tuned small models at one-tenth the price, with better performance and fully local deployment. Guillaume emphasized that many companies underestimate the gains from fine-tuning: “Using a closed model instead of fine-tuning on your own data is a sad thing, because your competitors are using the same model.”
Leanstral: formal mathematical proofs
Leanstral is a research project by a small Mistral team centered on Lean, a language for formal mathematical proofs. Guillaume explained its distinctive value: most RL training relies on verifiable rewards and suits problems whose answers can be checked, such as competition questions with numerical answers, but much real mathematical reasoning, such as “prove function f is continuous,” cannot be verified simply. Formal proof systems such as Lean and Coq naturally solve this: if the code compiles in Lean, the proof is correct, with no ambiguity and no judge model required. He predicts that as coding agents improve, software formal verification, currently an extremely niche field used mainly in safety-critical industries such as aviation and robotics, will expand substantially. Formal proofs also have an interesting side effect for agent architecture: when proving complex theorems, models automatically break them into lemmas and, to some extent, use subagents to verify them in parallel. Even if one lemma fails verification, others still provide partial reward signals, stabilizing training.
Mistral Small and multimodal integration
Guillaume introduced Mistral Small, a mixture-of-experts model with 6B active parameters and a 256k context window, as their first attempt to combine instruction following, coding, reasoning, and vision in one unified model. Previously, separate models from different teams carried these capabilities. Small is the first complete integration, with the larger Mistral 4 set to build on it. He clarified a misconception: for tasks such as pure ASR, or speech recognition, a model as large as Mistral Small is unnecessary; a dedicated 3B TTS model has overwhelming cost and efficiency advantages. General large models and specialized small models will coexist over the long term.
On hiring, Mistral is expanding in Paris, London, Palo Alto, Warsaw, Zurich, New York, and elsewhere, seeking applied scientists familiar with model training and RL pipelines, and forward deployed engineers who can work deeply on-site with customers to solve specific problems.