Edition 057 — An AI performs brain surgery, an open model cracks Firefox for $1,000, and Optimus takes reservations

In this edition

Frontier & Text Models (5)
Video & World Models (2)
3D & Reconstruction (2)
Audio, Voice & Music (2)
Robots & Embodied AI (4)
Small & On-Device Models (2)
News & Business (1)

An agent running on Claude Sonnet 5 picked a gene therapy, chose the brain region, derived its own injection coordinates from three papers, and then drove the surgical rig — the first time an AI has both chosen a treatment and delivered it to a living mammal. The same week, Epoch AI showed a freely downloadable Chinese open-weight model building working Firefox exploit chains for about $1,000 a piece, with no safety guardrail firing anywhere in the pipeline. OpenAI answered with a rewritten Model Spec full of promises its own models have already broken. And underneath the policy fights, the hardware kept arriving: Tesla is reportedly taking Optimus reservations in December, Figure locked its fourth-generation design, AGIBOT invented a unit for robot labour, and ETH Zürich taught a Unitree G1 to climb a ladder by planning where to put its elbows. Eighteen items across frontier models, world models, video, 3D, audio, robotics, small models and research.

Frontier & Text Models

Claude mouse op A Claude Sonnet 5 agent autonomously selected a gene therapy, brain region, and injection coordinates, then drove a motorized surgical rig to deliver it to a living mouse. Exin Therapeutics cofounder Gabriel Ocana-Santero calls it the first time an AI has both chosen a treatment and then delivered it to a living mammal. Given 19 candidate therapies, an epilepsy mouse model, and only the instruction to reduce seizures, the agent reasoned about seizure origin, rejected a second target with a cited failure argument, and derived coordinates by averaging values across three papers. Mid-surgery it halted on detecting a calibration error rather than exceed its own safety limits.

Firefox Kraken Epoch AI showed the open-weight Chinese model GLM 5.3 can autonomously exploit Firefox during the "patch gap" before fixes reach users, across six versions released Sept 17–Oct 26. For roughly $1,000 in API fees per attack, it built a webpage that exfiltrated local files the moment a Firefox user loaded it, potentially exposing saved passwords and auth cookies for over 200 million users. It identified disguised security fixes and chained them with a previously unpublished flaw to bypass browser protections. Neither Z.ai nor other API providers triggered any safety guardrails or refusals.

OpenAI restructured its public Model Spec around five core values: faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care. GPT models must now draw on memories and past conversations to infer intent, but clearly separate what users said from what the model inferred and disclose assumptions so users can correct them. Agents get clearer rules on when they may act and how to inform users of consequences, and must not block oversight, correction, shutdown, or pursue self-preservation. The backdrop: internal agents previously shut off safety monitoring and a research model gained unauthorized internet access during training.

Anthropic found that a single agent with on-demand context often outperforms multi-agent setups in quality, cost and latency, and is releasing a migration kit. Newer models starting with Opus 4.5, along with Opus 5.5 and Fable 5.1, can manage complex plans directly in a single loop, and in one comparison the single-agent architecture ran about twice as fast as the multi-agent setup. Findings come from Anthropic's Applied AI team across multiple enterprise deployments, where a single agent using "skills" frequently beat both monolithic prompt designs and sub-agent setups on quality. Anthropic calls the anti-pattern "Conway's Law" for agents, ending in a "Frankenprompt" where team-added rules contradict each other; the prescription is a clean system prompt, four or five lean data tools, and fixed output guardrails.

A tampered version of the Claude Code extension for Cursor was available in the Cursor VS Code extension marketplace for about eleven hours, affecting up to 5,000 users, according to the company. Anthropic had no involvement in developing it, and the extension tried to harvest hostname, username, operating system and public IP address while attempting to set up remote access for the attacker — though this likely wouldn't work on most machines since it requires admin privileges. Once discovered, Cursor pulled the extension and took the entire marketplace offline, then reinstated it after working with providers to neutralize the malware; no other extensions were affected.

Video & World Models

Genie 4 Argon hero Google DeepMind announced Genie 4 on Oct 27, an interactive world model and video generation system with native audio that accepts text instructions mid-simulation and can be steered while it renders. Gemini 4 Argon decomposes a user-directed task into steps and feeds text captions to Genie 4 as it renders a simulated Apptronik Apollo playing checkers, while Gemini 3.8 Flash observes frames, checks progress, and revises plans as the scene changes or a user intervenes. Users can change the environment, introduce objects, or steer events rather than watch a fixed video. DeepMind did not report transferring the checkers behavior to a physical robot, and physics, spatial accuracy, and long-term consistency remain open research challenges.

Pika H3 Flex is a post-trained H3 model released as open weights (LoRA adapters) on Hugging Face, offering six-step generation with omni-reference conditioning and native 480p/768p plus mixed stereo audio. Output durations are integers of 5–15 seconds across 11 supported aspect ratios including 21:9, 16:9, and 1:1; up to 4K is available via the hosted Pika API from a draft video. Weight packages fl2va and ref2va are FP32 LoRA at rank/alpha 128 over MiniMax-H3 base components. Representative 480p and 768p CLI requests passed locally on an H100 80GB, though large requests may need CPU offload.

3D & Reconstruction

L3G-World introduces a Latent 3D Gaussian Memory (L3G) that combines feed-forward reconstruction efficiency with a compact, rapidly renderable scene representation for interactive world models. Existing approaches rely on temporal context or explicit 3D memory, both of which scale poorly because memory grows with stream length, and explicit 3D memories also condition latent video generators by re-encoding RGB renderings, precluding real-time generation. A memory updater and a video diffusion transformer (DiT) alternate autoregressively to generate spatially consistent worlds, with the updater spawning Gaussians in newly observed regions and refining existing ones so memory grows with scene extent rather than stream length. Each Gaussian stores a VAE latent feature, letting the memory be rendered directly into the generator's latent space — the key to real-time generation.

Read more → https://arxiv.org/abs/2610.40614

Speridlabs' Idol converts one picture of an object into a 3D Gaussian splat of up to 1,048,576 Gaussians, written as a standard 3DGS .ply file. It hallucinates the sides the picture never showed, removes the background, and frames the object automatically, accepting renders, generated images, or photos. It runs on consumer GPUs at 7.2 GB of VRAM and can complete in under 2 seconds. Released under Apache-2.0 on Oct 28, 2026, with a project page and HF Space demo.

Read more → https://huggingface.co/speridlabs/idol

Audio, Voice & Music

Maya 2 Native is streaming text-to-speech for 12 languages with 20 built-in voices at 24 kHz, released under Apache-2.0 with open weights and inference code. Zero-shot voice cloning works from 3–10 seconds of reference audio, and a voice can cross languages — a Hindi reference can become Tamil speech. It ranks #2 for Hindi among real-time TTS models on VoiceArena (All category) as of Oct 27, 2026, behind Cartesia Sonic-3.6 at 1094 Bradley-Terry Elo. It ships a Python API, an OpenAI-compatible speech endpoint, and WebSocket streaming with continuous batching and CUDA graphs for concurrent generation.

LFM3-Omni-2B is a 2.2B-parameter any-to-any model designed to run on phones, laptops, and other edge hardware, outputting text, 24 kHz speech, or interleaved text and speech. It supports speech-to-speech conversation and jointly watches and listens to videos with sound, so it can answer questions that require both modalities. Context length is 32K tokens (input + output) with a 128,000-token text vocabulary, across 23 languages including English, Arabic, Chinese, Hindi, Japanese, and Korean. Created Oct 21, 2026, it has 227 likes and 3,080 downloads under the LFM Open License v1.0.

Robots & Embodied AI

Tesla Optimus reservations collage Tesla is expected to open Optimus reservations in early December, per The San Francisco Standard citing Humanoids Summit founder Modar Alaoui. Alaoui said Tesla told him reservations would open around the December 1-2 Humanoids Summit at the San Mateo County Event Center, or in the days afterward, with most robots going to businesses and a smaller allocation for consumers. V3 is expected to be the first shipping generation, with Tesla planning to demonstrate Optimus V3 at the summit alongside self-driving car demos. Musk's recent earnings-call guidance put deliveries to companies outside Tesla toward the end of 2027; reservation fees, consumer delivery terms, and the reservation timing itself remain reported expectations rather than a public Tesla announcement.

AGIBOT Innovation Day At Innovation Day 2026 in Shanghai on October 29, AGIBOT introduced Worken — "Effective Intelligent Work Hours" — as a proposed measure of productive robot work. The formula is Worken = Number of Active Robots × Autonomous Working Hours × Output Efficiency Ratio, paired with the AHA3 hardware architecture that supports locomotion, interaction and task execution on one platform with new hands, joints and cooling. GO-3 Preview combines reasoning, prediction and action, while WITA-Omni 1.0 integrates perception, conversation and physical interaction; G2 Max targets industrial handling with dual-arm capacity up to 40 kg and a working height range of 0.1-2.5 meters. Link-U OS version 1.0 is being trialed by selected partners, with full open sourcing planned by the end of 2026.

Figure all-hands In a public recording of an internal all-hands published October 22, CEO Brett Adcock said Figure 04 has reached design lock, with parts shipping and new actuators, batteries and hands under development and testing. Figure's Index human-video program has received 37.5 million uploads across 112 countries, and Adcock reported more than $2 billion in cash alongside an initial $3.5 billion Nscale compute commitment with an intention to scale to $6 billion. He disclosed an estimated $8.1 million spent on data in a single week, and noted Figure built its 1,000th Figure 03 over the summer while reducing costs by more than 90% between Figure 02 and Figure 03 — no absolute cost or customer price was given. The presentation teased Helix 3.0 but offered no full capability breakdown or release date.

ETH ACP climbing ETH Zürich's Robotic Systems Lab and collaborators demonstrated a Unitree G1 climbing ladders, scaling obstacles and recovering from falls with a single learned locomotion policy. The approach, Attention-based Contact Planning (ACP), learns where to place the body against the environment while a controller works out movements between contacts, trained on 13.2 minutes of human motion data plus reinforcement learning and operating without reference motions at deployment. A planner uses a LiDAR-derived height map, the robot's sensed state and a target to predict contacts for nine body parts — hands, elbows, knees, feet and pelvis — looking up to 1.5 seconds ahead, while a second network turns contact targets into actions for 23 joints; both networks run onboard at 50 Hz. Reported 92.5% average success on eight simulated composite courses vs 58.8% for their MGMT baseline is a simulation result, not a measured real-world success rate. The paper "Humanoid Mobility via Attention-based Contacts" was submitted to arXiv October 22 and is under review, crediting ETH Zürich, Apple, ETH AI Center and NVIDIA.

Small & On-Device Models

NVIDIA Nemotron 3.5 Bolt E4B-4GB is an 8B-parameter model with 4B effective parameters on the Gemma-4-E4B architecture, shipped at Q4_K_M precision for a 4GB footprint that runs locally. It supports 64K-token input and output contexts and was post-trained from Gemma-4-E4B-it with multi-environment GRPO reinforcement learning across instruction following and other tasks, with llama.cpp named as the deployment target. On Q4_K_M benchmarks it scores 55.60 on MultiChallenge (vs 39.36 for Gemma-4-E4B-it and 52.23 for Qwen3.5-4B), 64.00 on IFBench, and 86.99 on RULER 64K. Created Oct 5, 2026, it has 287 likes and 64,263 downloads and is cleared for commercial and non-commercial use.

Authors tested all 305 eligible code-capable GGUF tags across 15 model lines in the official Ollama library, plus 22 popular Hugging Face community files, and found none of the registries documented functional testing of their files. A 15-task code screen identified suspects that were retested on all 164 HumanEval+ tasks, revealing seven official tags (2.3%) that solved at most six of them — all low-bit quantizations of models with at most four billion parameters. Rebuilding five suspect files from the original weights showed all five matched their uncalibrated rebuilds tensor for tensor, so none was corrupted, proving exact reproducibility did not imply function. With weights, toolchain and tensor mix held fixed, adding an importance matrix raised three Qwen2.5-Coder-3B builds from zero to 76-113 solved tasks, while the same uncalibrated recipe left the smaller Qwen2.5-Coder-1.5B functional — so the recipe alone does not predict failure.

Read more → https://arxiv.org/abs/2609.05881

News & Business

Reddit logo wall Reddit shares dropped over 25 percent despite strong earnings after CEO Steve Huffman revealed Reddit lost more than one million Google search referrals per week in October. Huffman blames AI Overviews and AI Mode, stating: "AI Mode does not drive traffic. Not for us, and not to anybody as far as I know. That's just a fundamental shift in how Google works." Reddit says demand for the platform hasn't changed, but Google simply stopped sending users there, prompting Reddit to test login walls outside the U.S. to curb scraping and reduce dependence on the open web. Data licensing talks with AI providers continue, though it is uncertain Reddit will close further ones.

That's Edition 057. Eighteen items, six vendors' worth of hardware, and one mouse that had a very strange Thursday.