AI News Brief Roundup: 2026-11-10 Afternoon

AI News Brief: Afternoon roundup, 2026-11-10

Summary

The clearest thread in this window is AI agents running into limits from the platforms they touch. eBay appears to have blocked third-party agents, LexisNexis is building detection for them, and Rescale is putting agents to work on reactor engineering. Alongside this, money keeps flowing into AI infrastructure and data work, with Mercor's $20 billion valuation the largest figure, and Google is pushing its own chips and Gemini 4 Carbon as alternatives to Nvidia and rival models. A large share of the posts are arXiv papers on small models, local deployment, and evaluation, which show where the research effort is heading.

1. Mercor announced a $750 million Series D at a $20 billion valuation on November 10, with NVIDIA, Salesforce, and Zoom among the participants. The round included an employee tender, and the funds are earmarked to expand its AI training and evaluation work.
2. FormAssembly unveiled a platform overhaul, built over seven months, that moves beyond online forms into AI-driven business process automation. It combines form creation, workflow automation, integrations, and compliance management in one product.
3. An MIT Technology Review piece looks at Lubrizol, a specialty chemicals maker, and its push to let researchers test ideas faster with AI, cloud computing, and advanced modeling. The main obstacle it identifies is connecting these tools to reliable data and scientific expertise.
4. LoMASA is a research system that runs a multi-agent sales assistant entirely on one desktop, settling 38% of quoting decisions automatically at accuracy matching a rule baseline. It also made fewer commitments that later fell below the margin floor, at the cost of more manager escalations.
5. A reader told The Verge that eBay appears to block AI agents, including Meta's Muse, from browsing its listings. The report points to a growing barrier between consumer AI agents and major marketplaces.
6. The Import AI 480 newsletter covers self-sustaining space economies, progress in AI for science, and Meta's work on robotic hands for a superintelligence. It also highlights a startup that converted a quantum conjecture into a theorem.
7. This arXiv paper studies when separately fine-tuned models from a shared checkpoint can be merged without losing each other's gains. Its spectral selection method beats quality-only and cosine-based ranking on joint success, though it still shows a 17.1% regression rate.
8. Zyphra released ZONOS3-Mini, an Apache 2.0 text-to-speech model with 800M total and 100M active parameters. It clones voices, adjusts delivery word by word, and runs in real time in the browser.
9. GARP is a platform for real-time generative game characters that runs on local hardware using a 3B open-weight model inside the Godot engine. Its agents have emotions, Big Five personality axes, needs, and relationships, and passed 25 of 27 test runs.
10. Rescale introduced a framework for customizing engineering AI agents, and Kairos Power and Oak Ridge National Laboratory are using it on Hermes 2, a 50-megawatt fluoride-salt-cooled reactor. The agents handle thermal, flow, and shielding simulations along with the associated reporting.
11. This paper resolves a long-standing question in HTN planning by compiling SHOP-style method preconditions into partially ordered planning. It needs only two additional facts, rather than the auxiliary-action approach used by HDDL.
12. LexisNexis Risk Solutions is bringing Know Your Agent intelligence to government agencies through ThreatMetrix and BehavioSec, so they can distinguish humans, bots, and AI agents. The company reports 750% growth in agentic traffic across global digital transactions in the first half of 2026.
13. This paper finds that self-consistency voting on small language models often wastes compute and can flip correct answers to wrong ones. A gate using greedy-pass signals skipped 45 to 99% of voting calls, with point-estimate accuracy retention of 98.1 to 98.8%.
14. Fingerprint launched Fraud Score, a predictive signal that combines device intelligence with models trained on each customer's own fraud outcomes. In one head-to-head test it caught nearly 50% more fraud than the customer's in-house model without blocking more legitimate users.
15. The Hugging Face blog post "Your Experts Have Partners. There's HOPE for Keeping Them Together" has no text preview in the feed, so only its headline and source can be confirmed here.
16. Google expanded its AI hardware lineup with two specialized TPUs, one aimed at confidential model training and another offering more memory capacity. The move extends its chip portfolio to more use cases.
17. Google released Gemini 4 Carbon to the general public after a two-week preview with cybersecurity defenders. Google calls it its strongest model, says it can compete with top systems from Anthropic and OpenAI, and includes an agent "swarm" feature.
18. KDnuggets published a beginner's guide to five Python scripts for automating repetitive tasks such as organizing files and processing images. It is a practical tutorial rather than a news story.
19. This paper shows that AI-text detectors tuned on long passages become unreliable on short ones, and proposes length-binned calibration that holds the false-positive rate near 5% from 10 to 200 tokens. It also abstains where no detector is informative enough, which covers most short cases.
20. Google is expanding its TPU 8 family with the TPU 8tx, which keeps sensitive data encrypted during processing, and the TPU 8i Max, which doubles compute and memory by putting two chips on one package. With nearly 100 external customers, Google is positioning its chips as an alternative to Nvidia GPUs.
21. This paper argues that human-AI complementarity has been measured against ground truth labels, which fail for tasks like medical treatment recommendations. It proposes context-dependent adequacy labels built from a two-step, multi-criteria scoring method.
22. Google DeepMind's announcement that Gemini 4 Carbon is now available has no text preview in the feed, so only the headline is confirmed here.
23. This paper uses AI-guided refinement of partially formalized Lean proofs to improve bounds on analytic constants. It sets a new explicit upper bound for Grothendieck's constant, beating Krivine's 1977 result, and raises the lower bound for Korenblum's constant to 0.356565.
24. PuppyGraph raised a $10 million Series A led by Particle Future, with defy.vc participating, bringing its total funding to $15 million. Its annual recurring revenue grew sevenfold over the past year, with customers including AMD, Palo Alto Networks, and Coinbase.
25. Lightfoot is a graph-based retrieval system where deleting a document is a plain database operation with no model calls. On MuSiQue, its index is 26 to 31% cheaper than HippoRAG 2, while LightRAG took 2,187 model calls and 3.7 hours to delete the same documents.

Summaries written with AI from the linked source articles.

Sources in this roundup

Most-mentioned keywords

Sources

1. Mercor Raises $750M Series D at $20B Valuation to Expand AI Training
2. Exclusive: FormAssembly overhauls its platform to expand into AI-powered process automation
3. Innovating with AI in the special chemicals industry
4. LoMASA: A Local Multi-Agent Sales Assistant with Confidence-Gated Delegation for B2B Quoting
5. eBay appears to have shut the door on third-party AI agents
6. Import AI 480: Self-sustaining space economies; AI for science progress, and Meta builds perfect hands for a superintelligence
7. When Are Nearby Experts Composable? Cross-Task Model Merging in Neural Thickets
8. Zyphra Releases ZONOS3-Mini: Open-Source TTS With 100M Active Parameters
9. GARP: Real-Time Generative Agents on Local Hardware in a Game Engine
10. Rescale Puts Engineering Agents to Work on Hermes 2 Reactor Design
11. But It Was True When I Checked: Compiling SHOP-Style Method Preconditions into Partially-Ordered HTN Planning
12. LexisNexis Risk Solutions Brings AI-Agent Detection to Government
13. When Self-Consistency Hurts: The Hidden Cost of Majority Voting in Small Language Models
14. Fingerprint Launches Fraud Score, Reporting Nearly 50% More Fraud Detected in Customer Test
15. Your Experts Have Partners. There's HOPE for Keeping Them Together
16. Google broadens its AI hardware portfolio with two new specialized chips
17. Google releases Gemini 4 Carbon with agent ‘swarm’ feature
18. 5 Useful Python Automation Scripts for Beginners
19. How Short Is Too Short? Length-Conditional Calibration and Abstention for Detecting AI-Generated Text
20. Google ships four TPU variants in one year as it ramps up competition with Nvidia
21. Complementarity without ground truth?
22. Gemini 4 Carbon is now available
23. Improving Analytic Constants via Interpretable AI-Guided Lean Proof Refinement
24. PuppyGraph Secures $10M Series A After Sevenfold ARR Growth
25. Lightfoot: Graph Retrieval-Augmented Generation with Exact, Model-Free Document Deletion