Sep 2 · AI Daily

According to the WSJ, Gemini 3.8 Flash, codenamed Skimaki, could launch as soon as tomorrow. Google employees reportedly preferred it over Opus for coding in side-by-side tests. The article also mentions Google's internal coding tool Jetski, increased investment in RL by Google, and a look ahead to Gemini 4.

Dex Horthy introduced the /show-me skill in HumanLayer, aimed at making pull requests written by agents as easy to review as possible. Citing @notgiorgi, he noted that while agents know what acquire-use-release means, they implement it according to existing repository patterns. An included pseudo-code image illustrates a runQuery resource cleanup logic change and the potential error-catching differences it might mask.

YC-backed Nori Robotics introduced Nori A3, a low-cost humanoid robot built for developers, priced at $1,688 with shipping set for fall 2026 and assembly in San Francisco. Hardware features include 7+1 DOF per arm, a 1.5kg payload, 12-meter range Lidar, 4× 720p RGB cameras, and 6 to 8 hours of battery life. The accompanying Nori Lab laptop app lets users train the robot at home and upload skills to the Skills Marketplace to share.

Horizon Robotics announced that the Journey 6M has hit a milestone for large-scale mass production and deployment in vehicles. As of the end of August 2026, it has been adopted by over 20 automakers across more than 70 mass-produced models, with 7 automotive brands using the chip to roll out urban driver-assistance platforms. The Journey 6M delivers 128 TOPS of compute, supports single-model end-to-end urban driver assistance, and runs on air cooling with a single-chip setup to meet mass-production requirements for internal combustion engine vehicles.

Stryker announced that Duke University Medical Center completed the first hip arthroscopy using Apple Vision Pro paired with the SportSuite Vision app. The app displays arthroscopic video and CT imaging inside the headset, reducing the need for surgeons to repeatedly turn their neck to look at screens. Lead surgeon Chad Mather III stated that it allows critical information to be positioned within the sterile field according to individual workflows.

Claude Code released v2.1.258, fixing a regression introduced in 2.1.255 that prevented startup on macOS 12 (Monterey). It also fixes an issue where remote and scheduled sessions reported "user messages must have non-empty content" errors when permission approvals were re-sent and could not be applied.

Fable 5.1 features cache read costs 75% lower than Fable 5. As workloads lean more toward agents, the savings grow larger, with a reduction in one component driving roughly a 45% drop in cost for the entire workload. The author reposted an analysis noting that agents repeatedly revisit the same context, plans, tool outputs, and work states, meaning cheaper cache reads significantly lower ongoing inference overhead. Enterprises now face the financial question of how long an agent can work continuously before it makes financial sense.

Rohan Paul quoted and agreed with Ilya Sutskever's assessment: Neocloud networks have limited security, and rogue AI agents next might try to take over Neoclouds to run more copies of themselves, meaning network security needs a massive boost.

AI training data startup AfterQuery is rumored to have completed a financing round valuing it at $3.2 billion, just five months after raising a $30 million Series A at a $300 million valuation in April. Y Combinator partner Gustaf Alströmer called it the fastest company in the accelerator's history to go from founding to unicorn status. The two founders are 22 and 23 years old and part of the Y Combinator Winter 2025 batch.

Apple filed a new document in its ongoing lawsuit with OpenAI, disclosing alleged evidence obtained from a former employee's MacBook and continuing to push for expedited discovery. 9to5Mac reported that the filing contains what Apple calls striking evidence, though the material is only provided in summary form without specific details.

According to the Los Angeles Times, Google is contacting several major Hollywood studios to trade massive licensing fees for copyrighted content to train AI models. The generation rights for a single copyrighted character could be priced at around $40 million, with total deals potentially reaching billions of dollars and possibly including advertising revenue sharing.

MiniMax officially announced that, powered by vLLM-Omni and FastH3 (open-sourced by the FastVideo team @haoailab) with hardware support from @NVIDIAAI, a full 10.1-second MP4 video with synchronized audio was rendered in 8.7 seconds, generating faster than playback speed.

Rohan Paul cited test results showing that with guards disabled, Claude Mythos 5.1 built fully usable exploits in 245 out of 250 Firefox tests, achieving a 98% success rate, up from 52% for the previous flagship model six months ago.

Epoch AI data shows that since the launch of reasoning models, the Epoch Capabilities Index (ECI) frontier has progressed at a rate of 14 points per year, compared to 6 points per year in the pre-reasoning era. Charts show an adjusted R² of 0.98 for the reasoning model era with a 90% confidence interval of 12-17, compared to a 90% CI of 3-8 and an adjusted R² of 0.76 for the pre-reasoning era.

MiniMax published a post stating that if SOTA video models stay closed-source, we wouldn't see such widespread creation, adding that the community's projects have blown them away every day over the month since MiniMax H3 was open-sourced. In a reposted example, user @internetphysics built a generative video classroom using @fal's H3 Max, allowing users to input any concept to generate an animated explanatory video in seconds, queue up new videos while watching, and ask follow-up questions.

According to the WSJ, Gemini 3.8 Flash, codenamed Skimaki, will be released as early as Wednesday, with the author noting it is expected to match Opus 5. The report states that in head-to-head comparisons within Google's internal coding tool Jetski, some engineers preferred it over Anthropic's Opus model. The report also mentions that Google is ramping up investments in RL and offers a look ahead to Gemini 4.

Anthropic released Claude Fable 5.1 and Mythos 5.1 in response to customer criticisms regarding price, data retention, and overly strict safety limits. Fable 5.1 outperforms Fable 5, is generally about 25% cheaper, and cuts complex agent task costs by up to 45% due to cached data price cuts. It is now available across all platforms, while Mythos 5.1 is restricted to Project Glasswing participants.

World Labs released its next-generation world model, Atlas, pretrained from scratch to natively handle text, images, video, and 3D using a multimodal autoregressive diffusion transformer architecture where inputs form spatial context through 3D poses. The model supports camera-controlled generation (up to 1 minute of 1440p video), sparse-view 3D reconstruction, spatio-temporal simulation, and image generation. It outperforms more specialized models in camera control and 3D reconstruction tasks, and early access applications are now open, with the model set to power products like Marble in the future.

On September 1 local time, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. Both are built on the same base model with differing safety guardrail levels, and Mythos 5.1 is available exclusively to vetted cybersecurity and life sciences organizations.

The Hugging Face blog introduced BenchMIRT, a method based on multidimensional item response theory (IRT) to audit LLM benchmarks at the individual question level, using training data covering 100 LLMs, 16 benchmarks, and over 34,000 questions.