読書

Zoox has grounded its entire Atlanta test fleet after multiple safety drivers reported symptoms consistent with carbon monoxide, carbon dioxide, or hydrogen sulfide exposure inside the vehicles; one contractor filed an OSHA complaint.

Per OSHA letters viewed by TechCrunch, as many as 40 workers were possibly exposed, though Zoox says the number is lower. The company says it detected only CO2, will add CO2 monitors, and that OSHA has closed its inquiry. The source of the exposures remains unidentified.

The episode lands as Zoox expands into new cities to catch Waymo. Its purpose-built electric robotaxis are unaffected, but its test fleet is a rare one still running combustion engines.

Optical interconnect startup PicoJool announced a $27.5 million Series A round, bringing its total raised to $39.5 million.

The round was led by Socratic Partners with participation from Hudson River Trading. In November 2025, Playground Global — the venture firm of former Intel CEO Pat Gelsinger — led the company's $12 million seed round. No valuation was disclosed.

PicoJool builds VCSELs (vertical-cavity surface-emitting lasers), semiconductor laser diodes that move data inside data centers as light signals, aiming to replace the copper interconnects between AI chip clusters. Founder and CEO Al Yuen said the company has reached 200G per lane and is working with foundries including WIN Semiconductors to prepare mass production. The performance and efficiency claims are the company's own, with no customer orders or third-party tests yet; the new funds go first toward qualifying products with customers.

Ando came out of stealth on September 24 with a team messaging app designed for human employees and AI agents to use together, and announced $20 million in pre-seed and seed funding from investors including Accel, Index Ventures and Emergence.

The app gives agents their own identities and inboxes; they can browse channels, join conversations unprompted and notify colleagues on their own instead of waiting to be tagged. Founder Sara Du, who previously helped companies build MCP servers (a standard interface for connecting external tools to large models), says Slack and Teams treat agents as apps you install, while Ando treats them as team participants.

The amount and investor names are company-disclosed; valuation and round terms are not public. Slack has already turned its native bot into an agent and Microsoft has wired Copilot into Teams. Ando says it serves customers in software, real estate and finance across 15 countries, but most are small teams, and whether they stay is still unknown.

Basecamp Research announced a $140 million Series C led by S32, with more than a dozen investors joining, including Nvidia, NATO, and the Anthology Fund, the joint venture between Anthropic and Menlo Ventures.

The company's flagship product is EDEN, a 28-billion-parameter biology foundation model trained on the Trillion Gene Atlas, a dataset covering more than 100 billion genes; its initial focus is designing in vivo cell therapies.

This is an announced round only: the announcement discloses no valuation or terms. The capital will go toward accelerating drug development and expanding partnerships with pharmaceutical companies.

Swedish knowledge-management company Stravito on September 24 released an MCP server that lets clients query the consumer and market research stored in Stravito — reports, presentations, spreadsheets, audio and video — directly from AI tools such as ChatGPT, Claude and Copilot.

MCP is an open standard for connecting AI applications to external data and tools. The company says answers carry citations back to the underlying research, sometimes down to a spreadsheet cell, and inherit the access permissions already set in Stravito; the citations and permission controls are embedded in the server itself, not supplied by the protocol.

Founder and CEO Thor Olof Philogène said research can now sit alongside CRM and business-intelligence data in the same AI conversation. That is the company's own account, with no customer usage data yet.

Zhiyuan Robot says its 20,000th general-purpose embodied intelligence robot came off the line on September 24 and was delivered to Chimelong Group, with the count doubling from 10,000 in roughly six months.

Zhiyuan is a Chinese company building general-purpose embodied intelligence robots — humanoid machines that sense their surroundings and carry out tasks autonomously. Its stated production timeline: 6 prototypes in 2023, 1,000 units by January 2025, 10,000 by March 2026, and 20,000 on September 24. All figures are the company's own, relayed by tech outlet IT Home, and are not independently verified.

The deployment matters as much as the count: the same day, a themed park built with Chimelong opened in Hengqin, Zhuhai, with more than 300 Zhiyuan robots working regularly on guest interaction, guiding and shows. It is the first large-scale entry of humanoid robots into a high-traffic public venue, and day-to-day performance there is the observable test of what mass production actually delivered.

OpenAI announced on September 23 that the ChatGPT mobile app gains voice-based agentic features. Plus and Pro subscribers can use the Work tab on their phone to create documents, draft emails, or summarize Slack messages by voice.

This completes a rollout that began in July with the GPT-Live voice model and its desktop integration, where voice could already drive Work tab tasks on desktop. Conversations can now start on mobile and resume on desktop. Free and Go users are limited to plugins and connected apps.

The capabilities are OpenAI's own description, with no independent evaluation of task quality. For contrast, Anthropic recently merged its chat and workspace interfaces, while OpenAI keeps them separate.

Zhongke Leinao launched its compute-power-token integrated platform at the World Manufacturing Convention, placing compute, tokens and electricity under one scheduling system.

According to QbitAI, the platform follows the company's "1+3" architecture. The company itself reports that a cross-region scheduling test spanning Shanghai, Wuhu and Urumqi kept control response within 200 seconds, achieved a 100% migration success rate and 98% energy-consumption prediction accuracy. These figures are vendor-reported and independently unverified.

The same report notes the Wanjiang AI Technology Industrial Park in Urumqi opened on September 17, with its first fused computing center exceeding 1,000P already in operation. Whether the scheduling capability holds up in real business over time remains to be shown by operating data.

D-Robotics AI Lab has released Uranus, an open-source robot simulator built on a joint-trajectory-conditioned autoregressive diffusion model, with code, SDK and model weights published together.

The paper claims three capabilities: streaming rollout that consumes future joint trajectories frame by frame with no fixed horizon; 24 FPS after inference optimization; and a unified interface across robot embodiments and multi-view camera setups. All of these numbers are first-party, self-reported results with no independent reproduction yet.

One detail worth noting: version 2 of the paper was withdrawn and resubmitted as v3. The authors say the paper includes out-of-distribution evaluation and states current limitations. Teams doing robot data generation or policy evaluation can pull the weights and test the claims directly.

The Uni-LaDiR method self-reports a 7.3% relative gain over the strongest evaluated baselines across 11 vision-language benchmarks, and 6.1% across 2 robot manipulation suites, with no teacher observations used at inference.

The common prior approach concatenates modality-specific tokens; this method instead maps teacher reasoning steps via a unified encoder into shared latent-space tokens, and a diffusion model predicts the next block of thought tokens.

Note that all numbers are first-party evaluations with author-selected baselines and no independent replication; the preprint was submitted September 17 and revised to v2 on September 23.

Multi-party dialogue memory has a new method signal: SpeakerMem-R1 self-reports a 62.33% submission to EverMind-AI's public EverMemBench leaderboard, the highest reported, claiming to tackle both bottlenecks of who said what and reconstructing states across members and time. The method stores speaker-labeled verbatim messages and derived states in dual tracks, merging evidence by entity, event and time at query time.

Before this, no approach handled speaker attribution and state reconstruction together; the authors self-report binary accuracies of only 47.9%, 69.2% and 61.9% on GroupMemBench, SocialMemBench and EverMemBench.

In the authors' self-reported 305-question controlled test, RL lifted the SFT writer's mean accuracy from 57.38% to 68.20%; the leaderboard-best 62.33% is also the authors' own submission.

All numbers are first-party evaluations, not independently reproduced, and the sub-48% GroupMemBench score shows the problem is far from solved; the preprint material is dated 2026-09-22. Treat as a method signal to track.

Readers doing circuit analysis or safety evaluation now have a new idea to use: conditional co-ablation (CoAx), which after ablating a primary component set ranks candidates by growth in ablation effect to find backups the intervention activates — the authors report 0.941 ROC-AUC recovering known backup heads on GPT-2's IOI circuit.

Previously, recovering backup heads masked by primary components relied on intact-state attribution, whose strongest baseline reached only 0.815 ROC-AUC on GPT-2-small's IOI circuit and could not directly complete an incomplete circuit.

The authors' reported measurement: CoAx recovers the documented backup heads on that IOI circuit at 0.941 ROC-AUC, and adding the selected heads to the incomplete circuit cuts incompleteness from 0.75 to 0.21; they also report it beats matched random completions on 8 non-GPT-2 models across 6 architecture families.

All numbers are first-party evaluations, backup-head recovery is mainly validated on the known-answer IOI setting, and nothing has been independently reproduced; preprint arXiv 2607.01940 (v3, 23 Sep 2026).