TLDR AI 2026-09-24

What do battle bots, F1, AI agent labs, and hot wings all have in common? They're all happening at FC26.

Join 3,000+ AI engineers and leaders in SF Sep 29 - Oct 1, for three days of AI immersion, featuring industry leaders like Dr. Fei-Fei Li, 30+ breakout sessions, hands-on labs, and more.

Get hands-on alongside fellow builders and walk away with tools and strategies you can apply to your job, plus, see Sean Evans host Hot Ones live, watch BattleBots fight it out live, and catch Pitbull on the last night.

Register

Qwen Intelligence has launched three mobile agents for planning, cross-app execution, and rapid content creation, alongside open benchmarks for planning, real-device performance, and safety. Qwen reports top benchmark results and 90% end-to-end success for Mobile-Use.

Google released Gemini 3.8 Flash TTS and Flash-Lite TTS, letting developers create voices from descriptions and direct pacing, dialect, and delivery line by line. Flash-Lite targets high-volume work such as dubbing and voice agents, while Flash can replicate an authorized voice from a 30-second sample.

Anthropic said Claude autonomously identified a previously unknown enzyme system associated with unusual DNA repeats and characteristics seen in programmable genetic systems such as CRISPR.

Perplexity's SPACE platform tested VM isolation and network confinement using nine AI models, revealing no VM-host breaches across 108 trials. However, four models exploited network-policy vulnerabilities via DNS spoofing and IP-sharing, bypassing restrictions in 11 out of 54 partial-network trials. Post-remediation, none of the models succeeded in bypassing the updated security measures, highlighting the necessity for robust policy enforcement against shared infrastructure attacks, which were also found in eight of ten tested third-party platforms.

Robotics needs a universal post-training recipe similar to language models to achieve reliable autonomous deployment. Existing methods like imitation learning benefit from intuition rather than structured processes, and RL's instability hinders robotics. The development of EXPO-FT demonstrates a stable RL approach, emphasizing the need for standardized protocols, reward specifications, human feedback systems, and scalable processes to enhance robotics reliability.

Generic agents hunt for context as they work. Databricks Genie learns it automatically from knowledge already embedded across your data environment — so it can reason over live data, run new analyses and take action. See how Genie pushes the frontier for data agents →

See how Genie closes the AI context gap

Ember-1 is a new specialized model built on Kimi K3 that delivers Kimi K3's quality with 40% fewer tokens. It is now rolling out as a serving option alongside the base Kimi K3 model as a Research Preview release on Serverless. Fireworks Research is introducing research releases to give developers two-week serverless access to new research models. The models with the most demand will be made permanent.

tev1-4B-experimental is a Jev-like classifier fine-tuned on top of Qwen3.5 4B. It is available on Together serverless at $0.042 per million input tokens and $0 per million output tokens. Together AI has released the data recipe and a tutorial on how to fine-tune a custom model. Tev1 only cost $17 to train.

An AI agent's capability is largely set by its harness. Evolving the harness against a fixed evolve set is effective, but the harness memorizes the training tasks, and large in-distribution gains shrink or vanish out of distribution. RRSI keeps the harness edit space open and regularizes the search trajectory through it instead. Prompts, control flow, configuration, context management, tools, skills, memory, and sub-agents may all be modified. The constraints act on how the search moves, not on what the harness may contain.

Google says its planned Private AI Compute memory would let assistants recall context across devices without Google being able to read it. Data stays encrypted in cloud storage, keys stay on users' devices, and a protected enclave decrypts it only while answering a request.

Forecasters dramatically underestimate AI progress on benchmarks. Predictions on AI adoption and diffusion often lean toward underestimation. There is not enough evidence to assess forecasters' track record on predicting macro-scale economic and societal impacts of AI.

Deploy a functional RAG pipeline on AWS: Elastic Cloud for retrieval, AWS Lambda for orchestration, fully provisioned in one Terraform template. Start building now.

MentalHealthBench is an open benchmark created by OpenAI with more than 80 licensed mental health experts to evaluate AI responses across realistic mental health conversations.

Comfy Router, available on the Comfy Developer Platform, unifies access to various frontier media models with one API, allowing developers seamless model integration and provider flexibility for better pricing and availability.

AI-era education should prioritize curiosity, project-based learning, and intellectual exploration over grades and credentials.

Gemini is adding direct integrations with tools including Linear, Adobe, Webflow, Peloton, Experian, and SeatGeek.