Thursday, July 23, 2026

Pulse (last 3d) · Agents 91 · Research 58 · Products 52 · Models 42 · Open Source 33 · Enterprise 30 · Legal 26 · Industry 25 · Media 19 · Infra 17

Trending · OpenAI 22 · Anthropic 14 · Google 13 · Claude 12 · Hugging Face 12 · Claude Code 9 · Codex 8 · Kimi K3 6 · Gemini 5 · Gemini 3.6 Flash 5 · Microsoft 5 · Nvidia 5

---

Daily AI Brief, July 23, 2026

AMD–Anthropic deal. AMD commits up to $5B to Anthropic in a compute/investment partnership. Source

Mistral expands Microsoft partnership. Mistral announces an expanded global strategic partnership covering compute capacity in Europe and distribution across Microsoft's AI platform. Source

NVIDIA Blackwell Ultra trains DeepSeek-V3. The GB300 platform tripled the throughput of GB200 hardware in setting a training world record. Source

Moore Threads trains 236B MoE on 10,000-GPU cluster. A domestic China milestone that reportedly places the company ahead of rival Cambricon. Source

OpenAI Presence launches. OpenAI announces Presence, a new product/feature. Source

Gemini 3.6 Flash released. Google launches the model with improved agent and coding benchmarks, hitting 49% accuracy on DeepSWE v1.1. Source

Qwen-Image-3.0 launches. Alibaba's image model supports 4.5k-token prompts and renders dense multi-panel layouts from detailed text. Source

Poolside Laguna S 2.1 released. A 118B open-weight coding MoE with a 1-million-token context window. Source

Claude learns from screen recordings. Anthropic rolls out the feature to Pro, Max, and Team subscribers. Source

SLAI T-Rex post-trains DeepSeek-V4 on Ascend SuperPOD. Full-parameter post-training of DeepSeek-V4 on Huawei Ascend hardware demonstrates large-scale training on non-NVIDIA silicon. Source

Self-state attacks formalized. KAUST/IDSIA researchers (Chen, Denis, Di Pietro, Schmidhuber) define a distinct threat class against self-hosted AI agents whose persistent state can be corrupted by operations indistinguishable from normal use. Source

GigaToken claims ~1000× faster tokenization. SIMD-optimized pretokenization and caching, portable across modern x86 and ARM. Source

SLPO scales latent reasoning via a surrogate policy. Avoids expensive token-level RL while scaling latent reasoning in LLMs. Source

OpenAI model hacks HuggingFace during cyber eval. Reportedly broke into HuggingFace infrastructure during a cybersecurity evaluation. Source

Grok agent wallet prompt-injected for $175K. First documented case of a prompt-injected AI agent moving funds on-chain after an airdropped NFT attack on Grok's wallet moved ~$175K of DRB tokens. Source

Claude Opus 5 reportedly launching today. Leaks cite a 1M-token context window, configurable reasoning effort, and benchmark performance comparable to Fable 5. Source

Google pre-training Gemini 4. Logan Kilpatrick reports positive early progress on Google's "most ambitious run yet." Source