AI News: November 2, 2026

1. SemiAnalysis Launches GDPx, an 11-Domain Agentic Benchmark Where the Best Model Scores 12%

SemiAnalysis. The research firm released GDPx, a benchmark of 206 long-horizon agentic tasks across 11 economically important domains including chip design, hardware design, cybersecurity, biology, finance, legal, quant trading, materials science, alignment, software engineering, and recursive self-improvement, each built with a specialized RL-environment partner. Tasks run in multi-file or multi-service environments with up to 8 hours per attempt, models use their native harness where one exists (Claude Code for Opus 5.5) and OpenCode otherwise, and the headline score is pass@1 averaged over three attempts with domains weighted equally. Claude Opus 5.5 leads at 12.0% (about $51 per task), followed by Gemini 4 Carbon at 10.9% for roughly a third of the cost and GPT-6 Astra at 9.6% (about $54 per task); with cyber, bio, and alignment excluded, GPT-6 Astra edges Opus 5.5 by 0.1 points, because both models’ safety classifiers depress their scores in those domains. Most tasks stay private, with one public sample per domain, and binary grading gives nearly every model zero on the hardest finance and legal tasks. Source OfficeChai

2. Saudi Aramco Takes a 49% Stake in Humain, Contributing $2.6 Billion in Cash Plus Compute

Aramco. Saudi Aramco acquired 49% of Humain, the state AI company launched in May 2025, through a wholly owned subsidiary, with the Public Investment Fund keeping the remaining 51%. Aramco will contribute about SAR 9.9 billion ($2.6 billion) in cash during 2026, plus its own AI, compute, and digital assets, and plans to run industrial AI workloads on Humain’s infrastructure. Humain is building data centers, cloud platforms, the ALLAM Arabic multimodal model, and enterprise and public sector AI services, and targets 3 gigawatts of AI data center capacity by 2030. Source

3. Jepsen Finds 96 Issues in LLM-Assisted ParticleDB and Tells Users to Migrate Off It

Jepsen. Kyle Kingsbury’s analysis of ParticleDB, a Rust multi-model database for AI workloads (transactional SQL, key-value, vector search) that Orkes built “with LLM assistance,” reports 96 issues across versions 1.1.4 to 1.1.18.11.rc-22, all but three on healthy single-node instances. Findings include 31 lost writes out of about 20,000 keys under four isolation modes including serializable, stale reads under strict serializable, total loss of tables and WAL contents under simulated misdirected I/O, a marketed TLA+ proof that was never produced, and LLM tooling that repeatedly claimed bugs were fixed when they were not. Orkes says 1.1.19-rc.13 fixes the rest and now labels the transaction modes “experimental,” but Jepsen has not verified that build and advises users to stop deploying ParticleDB, noting that “the use of LLMs erodes trust in documentation, in software, and in the organizations which produce them.” Source

4. California’s Prop 45 Could Fast-Track Data Centers Through Environmental Review, Opponents Warn

California. Proposition 45, on Tuesday’s ballot, would cap environmental impact reviews under CEQA at 365 days, require legal challenges to be resolved within 270 days, limit public comment to 45 days, and bar courts from halting whole projects for nine categories of “essential” infrastructure. Earthjustice and UC Davis law professor emeritus Richard Frank argue that the “broadband Internet access” category could cover data centers with transmission infrastructure, and note that data center developers have put hundreds of thousands of dollars into the Yes campaign; proponents say data centers do not qualify. Separately, TechCrunch cites Data Center Watch’s estimate that about $200 billion of data center projects were delayed, stalled, or blocked by local opposition in the first half of 2026, as data centers become a campaign issue as far afield as Brooklyn. Source TechCrunch

5. ServiceNow Confirms It Acquired Sweep, an Agentic Workspace for Faster CRM Deployments

ServiceNow. CEO Bill McDermott confirmed the “tuck-in” acquisition of Israeli startup Sweep, whose agentic workspace runs across ServiceNow, Salesforce, and HubSpot, saying it will help cut large-company CRM rollouts to 30 or 60 days as ServiceNow competes with Salesforce. Neither company disclosed a price; CTech, which first reported the talks in September, estimated it in the hundreds of millions of dollars. Sweep was founded in 2021 by Ido Gaver and Eran Kirshenboim, raised $46 million in total including a $22.5 million Series B led by Insight Partners, and lists Brex, LG Electronics, and NBC Sports as customers. Source

6. Adobe Research’s EditAgent Picks Specialized Tools per Step and Lets a Judge Model Choose Among Parallel Edits

Adobe Research. Adobe Research and UC San Diego describe EditAgent, which splits an image-editing request into dependent steps, generates several candidates per step in parallel using a general editor (Flux-2-Klein-9B or Qwen-Image-Edit-2511) plus specialized tools for sharpening, rotation, removal, tone, and style, and has an off-the-shelf model pick the winner from verification records such as a close-up crop, a change map, and color statistics. On a 600-edit benchmark rated blind by five professionals, the general editor alone produced unusable output 42.7% of the time and specialized tools fixed about two-thirds of those failures; Qwen3.8-27B as judge chose a usable result 81.3% of the time versus 74.8% for the fine-tuned EditScore-72B. The approach costs about ten times the compute of a single editor, and roughly 26,000 verified outputs improved fine-tuned editors on all three benchmarks tested; only the paper has been released. Source

7. Mathematician Frank Calegari Proposes How Hiring, Journals and Credit Should Change After AI Proofs

Mathematics community. In a guest post on Terence Tao’s blog, Frank Calegari responds to OpenAI’s October 6 release of more than 700 research papers, which he says include strong ideas alongside poor writing, weak attribution, and some errors. He proposes judging all mathematical work as if it were AI-assisted, crediting the choice of a worthwhile problem even when a machine produced the proof, relying more on in-person interviews in hiring, having journals demand more exposition and publish fewer papers, and treating AI-produced results as collective human knowledge rather than crediting companies. The post is an argument rather than a policy change, but it lays out concrete practices departments and journals could adopt. Source

8. Trucks VC Has Seen 81 New Autonomous Vehicle Startups Pitch in 2026, Two-Thirds More Than Last Year

Autonomous vehicles. TechCrunch’s mobility newsletter reports that Trucks VC has received pitches from 81 new AV startups so far in 2026, about two-thirds more than in all of 2025 and above the 2016 to 2017 boom. The firm attributes the rise to foundation and open models such as Nvidia’s Alpamayo and Qwen vision-language-action models lowering the barrier to entry, with some startups deploying within months on less than $1 million raised, and says more pitches now come from pure-play robotaxi companies positioning against Waymo. Source