The week in AI

On the same day, David Silver, who led AlphaGo and AlphaZero at Google DeepMind, left to found Ineffable Intelligence in London. The company has raised $1.1 billion at a $5.1 billion valuation. Mistral released Mistral Medium 3.5 on 28 April, a 128B dense model with open weights that the company says scores 77.6% on SWE-Bench Verified.

Later in the week, Anthropic opened a public beta of Claude Security for Claude Enterprise customers on 30 April. xAI shipped Grok 4.3 the same day with video input and a lower API price.

Manus makes an AI agent sold as an app. Meta had agreed to buy it in a deal reported at $2 billion to $3 billion, and about 100 Manus staff had already moved to Meta's offices in Singapore. The NDRC order came after a months-long probe, according to TechCrunch.

The order is a first sign that Beijing treats moving a Chinese AI startup abroad as a technology export question. People in the industry call the practice "Singapore washing", and it means relocating a company's base to Singapore so that a foreign buyer can acquire it. Chief executive Xiao Hong and chief scientist Yichao Ji reportedly faced exit bans.

Chinese AI founders who planned a foreign sale through Singapore now have a case where the route was blocked. The week's records do not say how Meta would carry out the withdrawal or what happens to the staff who had already moved.

Silver spent more than a decade at DeepMind and led the work on AlphaGo and AlphaZero. AlphaZero learned chess, shogi and Go by playing against itself, with no human games in its training data.

Ineffable Intelligence is based in London, and Sequoia and Lightspeed backed the round. The company says it is building a "superlearner" that learns without human data, which carries the AlphaZero approach beyond board games. The records do not say what the first product or research result will be, or when one is expected.

Mistral Medium 3.5 is a single multimodal model with a 256K context window and a setting for how much reasoning effort it spends. Mistral also reports 91.4 on tau3-Telecom, an agent benchmark for customer support tasks in the telecom domain. The weights are under a modified MIT license, and API pricing is $1.50 per million input tokens and $7.50 per million output tokens.

The model now runs Vibe, Mistral's coding product, in place of Devstral 2. Vibe also gained remote agents. They run sandboxed coding sessions in the cloud in parallel, open pull requests and send a notification when the work is done, so a developer can hand off a task and come back to a pull request.

Claude Security scans code, checks whether each finding is a real vulnerability and writes proposed patches that a human reviews before anything is merged. Customers can use it from the Claude Platform or through partner tools.

The product appears to follow Claude Code Security, a limited research preview from February. It turns Anthropic's research on vulnerability finding into a product aimed at defenders. Anthropic has not published detection rates or false positive rates for the beta in the records available here.

• Grok 4.3 from xAI arrived on 30 April with a 1 million token context window and is the first xAI API model to accept video input, which secondary coverage says is up to five minutes at 1080p. API pricing is $1.25 per million input tokens and $2.50 per million output tokens for prompts under 200K tokens, roughly 40% to 58% below Grok 4.20, and the price doubles above that length.
• Qwen-Scope from Alibaba's Qwen team, released 29 April, is an open set of sparse autoencoders (SAEs) covering 14 SAE groups across seven Qwen3 and Qwen3.5 variants, both dense and mixture of experts. A sparse autoencoder breaks a model's internal activations into readable features, and the paper uses those features to steer language, concepts and preferences at inference without changing weights, and to check benchmark redundancy and guide data choices.
• ElevenMusic from ElevenLabs launched on 29 April as a consumer app for discovering, remixing and making music from a fully licensed model, with more than 4,000 independent artists who earn based on listener engagement. ElevenLabs says its voice marketplace, which works the same way, has paid creators more than $11 million.
• Seed 2.0 Lite from ByteDance Seed reportedly got an update on 28 April that added native audio input alongside video, image and text, which ByteDance describes as the first omni-modal understanding model in the Seed family.
• Ling-2.6-1T from Ant Group's inclusionAI, released 29 April, is a trillion-parameter model that answers without a separate thinking phase and mixes MLA and linear attention. Ant claims the best open-source scores on AIME26, SWE-bench Verified, BFCL-V4 and TAU2.
• AI co-clinician is a Google DeepMind research project, announced 30 April, for AI agents that work with patients under a physician's supervision. DeepMind reports zero critical errors in 97 of 98 primary care evidence queries.

• David Silver left Google DeepMind, where he led reinforcement learning, on 27 April to found Ineffable Intelligence in London.