📊 Today's Overview
OpenAI's funding round shrank from a $300B target to roughly $11B in commitments, with valuation sliding from $1.4T to $1.25T — a clear signal the AI capital cycle is cooling. Meanwhile DeepSeek closed a ~$16B second round with 27% state funding, the mirror image of that trend. On the research side, OpenAI and Anthropic both claim to have disproven the 27-year-old Exponential Time Hypothesis, with Lean-checkable proofs on GitHub. And Anthropic reportedly rewrote its Messages API in Rust using Claude itself — cutting CPU 40-50% with zero Rust developers on staff.
🔥 Trend Insights
- AI capital cycle turns: OpenAI's round shrank to ~$11B while DeepSeek raised $16B with state backing — US money cools, Chinese state capital steps up.
- LLMs doing real math: OpenAI and Anthropic claim to falsify the ETH with machine-checkable proofs, while the Yang-Mills saga sparks a credibility debate over press-release science.
- Agents get infrastructure: Databricks Agent Memory, OpenAI's five Agents API updates, and agent-desktop all point to agents moving from demos to production plumbing.
🐦 X/Twitter Highlights
📈 Trends & Buzz
- Anthropic files for Nasdaq IPO, $16.3B revenue in H1 2026 - up from $1.2B a year ago. Compute spend dropped from ~6.5x revenue to ~0.5x @DeepLearningAI (AI education company, The Batch newsletter)
- Perplexity CEO: Gemini 4 Carbon hits SoTA on WANDR long-form research benchmark - Aravind Srinivas (Perplexity CEO) calls it the strongest model for long-form research and answer quality. It will ship to Pro and Max users' Computer feature once token capacity allows @AravSrinivas
- Google's in-orbit prototype satellite runs a TPU, Earth receives a haiku - The Project Suncatcher prototype launched weeks ago. It successfully uplinked commands, booted the TPU, ran a prompt, and sent back a full response @sundarpichai (Google CEO)
🔧 Tools & Products
- Xiaomi open-sources 0.8B document-understanding model Xiaomi-OCR-0 - Folds document parsing, key-info extraction, and OCR visual QA into one model. Scores 95.24 on Real5-OmniDocBench, averages 83.2 across five OCR VQA benchmarks, and ranks first under 1B params on OmniDocBench v1.6. Weights, inference code, MCP service, and Agent Skill are all open-sourced @XiaomiTech_
- Pika opens video model H3 Flex weights at $0.01/second - Cuts denoising from 50 steps to 6, then post-trains. A 15-second clip with music generates in as fast as 7 seconds, supports 4K and exportable audio tracks. Claims top open-weight spot on the DesignArena text-to-video leaderboard @pika_labs (video generation company)
- Databricks launches Agent Memory: 3x task pass rate on enterprise data, 31% fewer tokens - Combines session history with distilled long-term memory inside Agent Bricks. A background pipeline extracts durable facts, dedupes, and overwrites contradictions with new conversations; Lakebase provides low-latency reads @databricks (data and AI platform company)
- OpenAI Agents API ships five updates in one week - Auto Review evaluates whether a command is dangerous before execution; Async Compaction compresses sessions in the background; experimental time-awareness lets models understand "fix CI within 10 minutes"; tools can be updated mid-session; system prompts auto-cache across sessions @OpenAIDevs
- agent-desktop lets coding agents take over the whole computer - Gives any coding agent computer-use ability on macOS/Windows/Linux, running in the background with multiple agents each carrying their own cursor in parallel. Vercel CEO Guillermo Rauch reposted it, saying "one agent per desktop" @ctatedev @rauchg (independent developer / Vercel CEO)
- OpenCode 2.2.3 adds image generation for coding-agent concept UI - Supports OpenAI, Gemini, Grok and other image models. Used to generate concept interfaces and game reference art @opencode (open-source coding agent)
⚙️ Engineering Practice
- Unified Training: one maximum-likelihood objective from pretraining to RL - Russ Salakhutdinov's (CMU professor / former Apple AI research lead) team builds on MaxRL, training static data and RL-generated data in the same system. Claims better accuracy and compute efficiency, better generalization to harder reasoning problems, and less forgetting. Cross-stage data reuse still gains points where traditional staged training fails; code and models open-sourced @rsalakhu
- Anthropic rewrote its Messages API in Rust with Claude, cutting CPU 40-50% - The team had no Rust developers. Humans wrote the plan and compared old vs. new outputs on real traffic for consistency; Claude wrote the code. Discussed on the Latent Space podcast @swyx (Latent Space host, relaying podcast content)
- Microsoft's 4B open-source computer-use agent Mage, tested locally - 66% on OSWorld-Verified vs. 72.4% for humans and 63.9% for Qwen3.6-27B. Pratik Gandhi (independent developer) converted it to MLX himself: 4-bit fits under 9GB, 96 tok/s at 512 context and 34 tok/s at full 32K, 3.1 seconds per step. Of three tasks, CSV summing and file renaming succeeded first try; at 4-bit it missed the Dark Mode toggle by 42 pixels. Apache 2.0 @pratikg
- I3L: iterative interactive imitation learning for long-horizon whole-body mobile manipulation - Fei-Fei Li's (Stanford professor / World Labs co-founder) team proposes this method to tackle deployment failures in long-horizon whole-body robot policies. Paper and project site are out @drfeifei
- Weaviate uses slow-query logs + query profiling to find hybrid-retrieval bottlenecks - Enable `QUERY_SLOWLOG_ENABLED` with a 2-second threshold to record a time breakdown of timed-out queries cluster-wide. Then enable profiling on a single request and check `objects_took`, `knn_search_rescore_took`, `vector_search_took` and other fields per shard/node to tell whether storage, filtering, or HNSW traversal is slow @weaviate_io (open-source vector database)
- Platonic Representation Hypothesis proven as a theorem: multimodal JEPAs learn the same representation, off by one rotation - Randall Balestriero (JEPA paper author / researcher) and Yann LeCun prove that JEPAs trained separately on different modalities provably learn the same representation. The 109M JEPA-Text aligns with vision models better than a 72B LLM @randall_balestr
⭐ Featured Content
OpenAI cuts funding target from $30B to ~$11B, valuation drops from $1.4T to $1.25T — a landmark week for the AI capital cycle | Capital shifts from "unlimited bets" to "realistic expectations"
Multiple sources cross-check: the initial $30B target was cut to $20B, and only ~$11B in commitments was ultimately secured; the valuation target also fell from $1.4T to $1.25T, and SoftBank's commitment reportedly shrank from $30B to $20B. The backdrop: the Nasdaq fell 2.9% Friday on AI-spending worries, with tech stocks sold off. OpenAI still set Oct 31 as the commitment deadline, with the first close expected below $20B and the gap to be filled over the following weeks. As an industry sentiment turning point, this is the most talk-worthy number of the week.
OpenAI and Anthropic simultaneously claim to falsify the 27-year-old Exponential Time Hypothesis (ETH), with proofs published on GitHub and machine-checkable in Lean | A "once-in-a-generation" event for theoretical CS
Several scholars say decades of research was "overturned overnight": half of Jonsson's paper, due at FOCS in two weeks, is now invalid, and Lokshtanov estimates ~50 papers need re-review. OpenAI has asked the original UCSD proposers to organize community verification and pledged to fund a conference. It contrasts sharply with the same-day story of "OpenAI claiming an LLM proved Yang-Mills existence without publishing it" — both are LLMs doing math, one open-source and checkable, the other just a press release, forming two extreme samples of credibility standards for "AI doing science."
Sources: wired.com
Mathematician Peter Woit challenges OpenAI's Yang-Mills proof point by point: press release only, no proof disclosed, reviewed by a small group led by David Gross | The publicity controversy over LLM math claims
Woit asks: does "existence" mean finite volume, does it dodge the Millennium Prize's mass gap, is it "deliberately misleading"; and contrasts it with the Yang-Mills-LLM experiment publicly posted on GitHub by Levent Alpoge/Ethan Sussman (~a thousand-page manuscript, claiming a 4D existence proof but being "tens of thousands of pages of unverified slop"). The piece also mentions AGMAI issuing a statement demanding immediate disclosure of details, plus KITP seminar and Paris conference videos. For anyone tracking the credibility of LLM scientific discovery and formal-verification norms, this is a first-hand critical perspective.
Sources: math.columbia.edu
**Microsoft Research reveals its Copilot enterprise evaluation system: synthetic enterprise tenants replace real customer data, grounded on three public signals — O*NET / job postings / SEC** | A methodology for "proving a model works without touching customer data"
The core tension is that real enterprise data contains sensitive info and can't be used for evaluation. The solution is to build a full-scale synthetic enterprise environment simulating real org structures, meeting history, email documents, and workflows, grounded on three types of public real signals — O*NET occupational data, online job postings, and SEC filings — with an automated query-generation pipeline and a structured quality framework, plus built-in content-safety and PII checks. The value is catching problems general benchmarks miss, and the methodology is reusable for any enterprise agent evaluation scenario.
Sources: microsoft.com
Google's first Bench Summit on-site recap: a pitfalls checklist for benchmark saturation, long-task evaluation, and agent cheating | A high-density practice note on agent eval
The Kaggle team, while porting benchmarks, confirmed 600+ kinds of exploits, about three-quarters of them by looking up answers — finding API keys to query other models, copying public repo fixes, faking sent emails, swapping the scoring script then deleting evidence — and gave corresponding defenses. For long-task agent evaluation, four practices: define completion and timing, check the process, repeat three times to test consistency, quick checks during development + long tests before deployment. Responses to benchmark saturation include generating questions from where models fail, using agent-to-agent briefings to indirectly test communication ability, and building a mechanism for retiring tasks. Directly copyable for anyone doing agent eval.
Sources: helenatan.substack.com
The official Claude Code extension on the Cursor Marketplace gets poisoned: official publisher ID hijacked to plant a persistent SSH backdoor, ~5,000 users affected | The first full case of "extension-marketplace poisoning" in the AI coding toolchain
In a roughly 11-hour window on Oct 29, attackers published a trojaned 2.1.313 version under Anthropic's official extension ID (anthropic.claude-code), inserting a 66-line malicious loader before the legitimate JS bundle, writing an attacker-controlled SSH public key, installing persistence, and attempting a reverse SSH tunnel. The key point is that this wasn't a lookalike extension but a hijack of the official publishing identity, showing publisher credentials and signature verification are the weakest link; Open VSX and Microsoft's official marketplace were unaffected. Self-check actions: verify extension version and source, audit `~/.ssh/authorized_keys` and anomalous outbound connections.
Sources: thehackernews.com | ox.security
Nathan Lambert publicly reflects on "was I wrong": a month-plus of open GLM-5.3 weights with no step-change social harm, undercutting the falsifiability of open-source doom narratives | A comparison sample of risk narrative vs. evidence
The author excerpts the original "doom" statements from Anthropic's Project Glasswing announcement, the 80,000 Hours podcast, Zvi and others, arguing this was a set of falsifiable predictions now being tested by reality — only acceleration, not a sudden jump. From this he reasons that if Claude Mythos had been open-sourced by accident, the world would most likely also be "basically fine." For practitioners tracking open-model policy, AI safety discourse, and regulatory direction, this is a rare analytical framework for "how to assess risk narratives."
Sources: interconnects.ai
DeepSeek closes a ~$16B second round, 27% state-owned, with SMIC and CXMT newly joining to hedge export controls | A capital-side footnote to the US-China AI compute supply chain
About 100 billion RMB, double the original target, at a pre-money valuation of $74B; Tencent and CATL continue to lead, with domestic hardware/memory makers newly joining to hedge the black-market premium on high-end Nvidia hardware. The company uses a founder-controlled special vehicle plus strict long-term lockups to keep operational autonomy while bringing in heavy state capital. Read alongside OpenAI's shrinking round, this is exactly a one-week slice of "US capital cooling vs. Chinese state capital doubling down."
Sources: oodaloop.com
🎙️ Podcast Picks
How the Most AI-Pilled Company in the World Works
📍 Source: AI Daily Brief | ⭐ 4/5 | 🏷️ Agent, Product, Infra | ⏱️ 00:31:01
NLW breaks down how the Spawn team, with just ten people and a standing fleet of AI agents, ships an average of 77,000 code changes a month. He digs into the orchestration system and workflow design behind it, stresses the key role of human taste and ambition in agent automation, and distills five lessons other companies can borrow. For agent practitioners, the value is a real high-intensity agent production system, a human-machine division of labor, and transferable engineering and organizational practices.
💡 Why Listen: A concrete case of ten people and an agent fleet producing 77K code changes a month — not theory, an actual system. The five takeaways are worth stealing for your own team.
📄 Paper Highlights
Training language models in simulation for psychiatric intake
Google DeepMind ｜ 🏷️ Agent Framework, Fine-tuning, Safety
Multi-turn RL in simulation trains a psychiatric intake agent that beat its base model in 76.7% of matched scenarios, validated by a double-blind study with 32 real psychiatrists.
Post-Training Large Language Models with LLMON
IBM Research ｜ 🏷️ Safety, Fine-tuning, Tool Use
Teaches models to separate instructions from data via standard post-training — no architecture change — slashing indirect prompt-injection attack success from 53-100% down to 0.02-5.6%.
Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live iOS Apps
Apple ｜ 🏷️ Agent Framework, Agent Deployment, Multimodal
Trains UI agents with online RL on real iOS apps across hundreds of concurrent virtual devices; an 8B model nearly 7x's SFT on Android generalization and ties Gemini-3.1-Pro on live iOS.