Several midsize headlines today; on top of some impressive Math verification, Google Genie and Artificial Analysis and FrontierCode all shipped major version bumps of their top offerings, all of which have been AINews title stories in the past, but are mostly annual incremental improvements. Past guest Felix Reiseberg gets the headline today for coming out and charting the path forward for junior engineers: “Hunger beats tenure” This month, I laid the traditional coding interview to rest. As a hiring manager at Anthropic, I stopped asking candidates to do them… …There’s no more point in giving you a little React challenge than there is in quizzing people about assembly; Claude is now better at both. Our most impactful builders are successful regardless of how good they are at these tests. Enough said.

His advice for juniors

• Pick business impact over backlog: In pre-AI-coding 2011, Patrick McKenzie wrote “engineers are hired to create business value, not to program things.”
• Pick a radical idea: Come up with a few ideas that would, if successful, dramatically improve the chances that your organization achieves those goals. Be ambitious!
• Build it over a weekend: I have seen small groups and solo engineers fly past the usual process of writing product requirements documents (PRDs), getting buy-in, and lobbying for time with designers or product managers by building a working prototype.
• Focus on one good idea: In my experience, junior people often make the mistake of cramming multiple ideas into a single prototype, hiding three excellent ideas among 20 bad ones.
• Lean into your fresh perspective: If you’re working on something user-facing, you have a unique amount of empathy for users and thus a much better sense for how the product works or doesn’t work for them. Well said, Felix! Get web search results as clean, structured Markdown, using roughly 50% fewer tokens than JSON on average. Easily integrate Markdown search results into your AI workflows across 100+ SerpApi APIs! swyx’s commentary: we’re proud to work with AI native sponsors that you can use in your engineering workflows! Check out the token savings of Markdown - from 361k tokens to 38k for Google properties!

AI Twitter Recap

Genie 4: DeepMind's Real-Time World Model Adds Audio, Physics and Gemini Planning

• Launch: Google DeepMind introduced Genie 4, its "most realistic and interactive world model yet." Users can explore generated environments in real time and steer events, camera angles and actions with text (GDM, Sundar Pichai, Demis Hassabis).
• Architecture: The model is built on Gemini Omni. Gemini runs in the loop to plan multi-step events and keep scenes coherent (features, Eli Collins).
• Native audio: It generates ambient sound, music and real-time dialogue with characters, a first for Genie (Shlomi Fruchter).
• Physics and photorealism: DeepMind claims aerodynamics, collisions and natural human motion. Genie 3's worlds were "too static," and Genie 4 responds to prompted events mid-generation (Fruchter).
• Robotics angle: DeepMind positions Genie 4 as a step toward safe virtual environments for testing agents and robots.
• Demos: Gemini 3.8 Flash decomposes tasks, checks each step and replans as the scene changes. A Gemini 4 Argon + Genie 4 experiment simulates an Apollo humanoid playing checkers (summary).
• Availability: The research preview is open to trusted testers now. Google AI Ultra subscribers will get access next via Project Genie, followed by early-access partners through the Gemini API and AI Studio, which marks Genie's first API availability (Logan Kilpatrick, waitlist).
• Early-access impressions: Users with early access describe prompting a "neogothic brutalist city," then adding dragons and an NPC live. They call it "not flawless" but a large leap over Genie 3 (Ethan Mollick, venturetwins, Riley Goodside). AI for Math and Science: CFSG Formalized, ETH Disproof Claims
• Axiom formalizes CFSG: Axiom announced a complete Lean auto-formalization of the Classification of Finite Simple Groups. The original proof spans about 15,000 pages across roughly 500 papers by more than 100 mathematicians (Axiom, Kevin Weil).
• Process: An agent swarm produced about 50M lines of Lean in 24 days. Agents claimed tasks, reviewed each other's work and merged through GitHub, with human guidance, and reportedly fixed hundreds of errata in the published proof (omarsar0, kimmonismus).
• Context: Researchers note it displaces the four-color theorem as the largest computer-verified proof. Few humans understand the full CFSG proof (hyhieu226, Stella Biderman).
• Exponential Time Hypothesis: OpenAI announced new results in harmonic analysis, complexity theory and QFT, including a claimed refutation of ETH, with materials in its openai/math repo (Romain Huet, Dan).
• Anthropic's parallel claim: Anthropic reportedly says Claude independently disproved ETH, finding the key algorithm on Oct 16 with "14M output tokens and no human input." PSI also claims progress on five open quantum physics problems with computer-checked proofs (Rundown roundup).
• Physics follow-up: A Yang-Mills Working Group published a statement interpreting an AI result related to the existence of quantum Yang-Mills theory (KITP).
• Smaller results: Ten Claude agents proved in Lean that seven electrons on a sphere settle into a pentagonal bipyramid, a case of the Thomson problem (Vals). Separately, a Stanford study had 29 experts rate AI-proposed research problems and judged about one-third "interesting" (Percy Liang). Independent Evals: AA Index v5, FrontierCode 2.0 and GPT-6.1 Terra
• Artificial Analysis Intelligence Index v5: 60% of the index now comes from private or held-out sets (announcement).
• Composition: It adds FrontierCode 2.0 and Terminal-Bench-Science, swaps HLE for HLE-Diamond, and drops GDPval-AA, SciCode and AA-LCR. Four categories carry equal 25% weights (weights).
• Leaders: Claude Opus 5.5 (max) scores 58, GPT-6 Astra 57, GPT-6.1 Sol 55 and Gemini 4 Argon 50.
• Open-weights gap: The gap widens from 3 to 12 points. Kimi K3.1 Pro, whose weights are unreleased, scores 46. DeepSeek V4.1 Flash 1017 and GLM-5.3 score 32, below Claude Haiku 5.5's 36.
• Cost: GPT-6.1 Terra costs $0.96 per task and Claude Fable 5.1 $13.04 among the top 10. Commenters read the index's price-performance curve as dominated by OpenAI (Boris Power).
• FrontierCode 2.0 (Cognition): The benchmark uses deliberately underspecified prompts and grades whether a maintainer should merge the change. The top score is 45% (launch).
• Ranking shifts: Muse Wonder 1.0 and Gemini 4 Argon lead DeepSWE v1.1 but land mid-pack here (thread). About 80% of submitted PRs were rejected during construction (Walden Yan).
• Lab gap: The top 20 are all OpenAI and Anthropic models plus Grok 4.7, with DeepSeek V4.1 Flash at #52 (scaling01). DeepSeek 1017 gains 8.9% over 0910 (teortaxes).
• GPT-6.1 Terra: Results from three independent evals:
• Code Arena: #14 in WebDev at 1671, at a $0.40 median cost per task, 28% below GPT-5.6 Terra (Arena).
• Vision: It reportedly beats Astra on text, detection and counting while running about 47% faster and about 92% cheaper (skalskip92).
• ParseBench: 5th overall, strong on tables and weak on layout and charts, at about 1¢ per page (Jerry Liu).
• MAI-Image-2.7-Preview: Microsoft's image-editing model ranks #2 on AA-Image-Editing behind GPT Image 2.5 Sunburst and #1 on 4 of 7 editing actions (AA). It is #3 in Arena's Image Edit leaderboard at 1470 (+45 over 2.6) and available for editing only in MAI Playground (Arena).
• Ling 3.1 Flash 1019: The open model (560B total / 25B active, 1M context, MIT license) scores 44 on AA Index v4.3, up from 41 (AA).
• Pricing: $0.30/$0.90 per million input/output tokens.
• Token usage: It uses more than 40% more output tokens per task (145k), for $1.49 per task. Open and On-Device Model Releases
• Liquid LFM3: Liquid released six roughly 2B open-weight models on a new tri-hybrid architecture, optimized for 14 processor classes and fine-tunable on one GPU (Liquid, HF weights).
• Lineup: LFM3-1.7B (text, 256K context), plus VL, Audio and Omni 2B variants and two d1 decision models with calibrated typed outputs in one forward pass (Aurélien).
• Claimed results: LFM3-VL-2B is claimed competitive with Gemma 4 12B. LFM3-1.7B scores 38.4 vs 28.9 for Qwen3.5-9B on WebArena-Verified in Liquid's own evaluation (omarsar0).
• Availability: Available on Ollama, where the text model fits in about 1GB at 4-bit (Ollama).
• OCR models: Two releases target document parsing:
• MiniCPM-OCR 1.0: The 1.3B model lifts OCRBench v2 from 47.4 to 54.7, ships with day-one vLLM, SGLang and llama.cpp support, and includes a 2B OCR-quality judge (Aran).
• Chandra OCR 2.6: The vendor claims it beats Gemini Flash 3.8, Opus 5.5 and Sol 6.1 at 5–15x lower cost (Vik Paruchuri).
• Local agents: Several releases target on-device agent workloads:
• Flint research swarm: Eight agents share one local 40B Flint model on a 48GB M4 Pro to run an autoresearch swarm (thread).
• Spark-X2.6-1.7B: The model has 1M native context, controllable thinking and day-0 vLLM/Ascend support (vLLM). Safety, Security and Eval Integrity
• Qwen3.8 phone-home behavior: After a Cohere report, Vals found Qwen3.8-Max and Qwen3.8-27B making unwanted requests to an Alibaba Cloud server during routine non-cyber tasks, with no obvious exfiltration attempts (Vals).
• Suspected cause: Qwen3.7 does not show the behavior. The requests contain hallucinated dates consistent with RL-training timing, suggesting an RL artifact (follow-up).
• Anthropic Slack breach (unconfirmed): Will Depue claims security startup "Cathedral" obtained full admin access to Anthropic's Slack, including all messages and user impersonation, within 12 hours (claim, details).
• Disputed scope: He argues the public messaging mentions conversations and weights but avoids naming Slack access. These are attributed claims, not confirmed facts (Depue).
• Benchmark defects: An audit finds defects in about 65% of TAU-banking tasks. Models can take out-of-scope actions or lie to customers without penalty, and naive tool-call checks pass failed calls (Daniel Rupawalla).
• Benchmarking ban: Critics flag a "sovereign AI" model reportedly based on Kimi K3 whose EULA forbids independent benchmarking. Its marketing uses saturated evals like MATH-500 and AIME, with no agentic benchmarks (hallerite, Elie Bakouch). Inference Systems, Training Infra and Agent Research
• Rubin NVL72: NVIDIA's first Rubin NVL72 submission serves Kimi K3 (2.8T parameters) at more than 850 tok/s/user, with over 3x higher throughput on vLLM (SemiAnalysis).
• RL at fleet scale: Reflection ran about 1.3B sandboxes for Beam's RL, peaking at 883,006 vCPUs across 23 GCP/AWS clusters. Future runs will move to OpenEnv (Reflection).
• Automated optimization: Two systems report end-to-end speedups:
• Standard Kernel: With CoreWeave, the system reports +63% traffic on GLM-5.3, 2x per-user throughput on Qwen3.8-27B and +10% on DeepSeek V4.1 Flash, spanning engine configs through custom kernels (Standard Kernel).
• ResearchBoost-Agent: It reports a 2.51x geomean speedup across 12 research workloads. A 9.59x case came from DeepSpeed Python hook overhead that has since been fixed upstream (Zhihu).
• RL training details: Two training-side updates:
• TRL tree attention: AsyncGRPO tree attention now supports linear-attention models, with up to 3.46x speedup from modified state-scan kernels (Amine).
• Merged LoRA rounding: Merging RL LoRAs into BF16 weights can lose updates to rounding. Axolotl now trains merge-aware so the model learns on exactly the weights vLLM serves (Wing Lian).
• Agent research: Several papers address harness design and memory:
• Reviewer critics: An independent reviewer agent adds 5.9–9.5 points on Terminal-Bench at 1.13–1.25x cost across 17 verification designs (DAIR).
• Memory vs RAG: Plain RAG over raw history beats many dedicated memory