Profiling a batch-size-1 evaluation of Inclusion AI's Ling-3.0-flash on NVIDIA DGX Spark showed decoding was memory-bound, with model weight reads driving execution time. The MoE computation and the output-layer projection that produces scores across the model stood out as the two key operations.
Luis, who has studied AI for about 10 years, argues that curiosity-driven play at the edge of one's knowledge, as with Feynman's wobbling plate, is the engine of real learning, and outlines eight "threshold concepts" that explain modern AI: learning from data, representations and embeddings, next-token prediction, attention, evals, post-training and RL fine-tuning, harnesses and agents, and recursive self-improvement. He cites Vygotsky's zone of proximal development, Meyer and Land's threshold concepts, and Bruner's spiral learning, and includes a study prompt usable with any LLM. The post ends with a paragraph promoting his AI coach, Zairai.
OpenAI announced on October 26, 2026 that its internal models produced three results, including a deterministic algorithm that counts satisfying assignments of k-CNF formulas and thereby refutes the Exponential Time Hypothesis, with the proof formalized in Lean and a working group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams set to verify it. The paper's gains are asymptotic and very slow: with all hidden constants set to 1, the general bound only beats the best known 3-SAT algorithm past roughly 540 million variables, and the Lean theorem proves a weaker bound that beats plain 2^n only when n has about 2,400 digits. The paper itself notes that P versus NP is untouched, and OpenAI says it seeks no academic credit and will fund a conference on the result.
Why it matters: Useful for readers tracking AI-generated mathematics because it separates the asymptotic claim from practical impact and details exactly what the Lean formalization does and does not establish.
Former Deepmind researcher Thore Graepel founded Metis Reasoning, which proposes applying AlphaGo's approach of pairing neural networks with systematic search to general reasoning, using an explicit "knowledge ledger" that records what is known and checks new entries against evidence. Metis says the setup should produce fewer hallucinations and knowledge that can be changed without retraining, though the white paper describes only a research program with no working system or benchmarks. Graepel, who worked on AlphaGo, AlphaZero and MuZero, is reportedly raising a low tens-of-millions round for Metis.
Mirendil released Fern 0.1, a from-scratch coding model with a context window of up to 10M tokens (about 23 MB of code), built on Mirendil-Base after a short round of post-training and available only through hosted early access. Mirendil says one researcher pretrained the base model in 3.5 months for under $10M, with AI agents writing over 99% of the code and training data drawn only from public sources. The base model scores 100% on RULER single-needle retrieval from 32K to 10M tokens, while Fern 0.1 scores 6.1% on Terminal Bench 4.0 and 73.4% on SWE-Bench Verified, trailing the frontier and large open models in its comparison table.
The vLLM Semantic Router team released d3-ultra, a 26.09B-parameter dense multimodal decision model under Apache-2.0 that does not generate text but reads an input once and returns probabilities for typed questions (choice, yes/no, score, box, track), with 131,072 tokens of context per question. It scored 87.2 on the team's private SR-Text test, the highest of six decision models compared, but trailed its sibling d3-turbo on SR-Box grounding (58.2% vs 60.0% over 10 domains). It was tested on one AMD Instinct MI325X with BF16 weights, where 255 questions about a photo took 2.16 s, and is available via Transformers, the vLLM Semantic Router Engine, a free API and a Hugging Face Space.
A link-blog post notes that OpenAI's GPT-6.1 Terra announcement emphasized visual reasoning and computer use, describing it as a fast, efficient model positioned between the cheaper Luna and the more expensive Sol, priced at $1/million input tokens and $5/million output. The post quotes OpenAI's Ted Sanders saying Terra is better than Astra and Sol at technical communication, similar on vision and computer use but faster, and a bit worse on most other things, and includes pelican test images across reasoning levels.
Featured coverage of this story: 《Introducing GPT-6.1 Terra》
OpenAI released GPT-6.1 Terra, priced at $1 per million input tokens and $5 per million output tokens—half the cost of GPT-5.6 Terra and below Claude Sonnet 5.5's $2/$10. Artificial Analysis reports it gains 6 Intelligence Index points over GPT-5.6 Terra at maximum reasoning effort while costing $0.40 versus $1.40 per task, and it solves 50.3% of Terminal-Bench 4.0 tasks versus 21.5% for its predecessor. It is available now for Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex and via the API, though OpenAI still recommends GPT-6.1 Sol and GPT-6 Astra for the most demanding coding tasks.
Featured coverage of this story: 《Introducing GPT-6.1 Terra》
In this Decoder interview, Kevin Roose discusses leaving The New York Times with Casey Newton to launch the video-first podcast Machine Gods, and his new book The AGI Chronicles, a reported history of the frontier labs racing to build artificial general intelligence. He describes how the transformer gave companies a tool for a long-held dream, how labs differ in how "AGI-pilled" they are (with Anthropic most and Google least), and how belief in AGI drives both bet-the-farm compute spending and a struggle to focus on business fundamentals.
This Import AI issue covers Resolution's autoformalization of the resolution of singularities in Lean, a roughly one-week effort costing an estimated $135,000 in tokens and producing about 600k lines of Lean, framed as groundwork for automated research into loss landscape geometry and generalization. It also reports Claude Opus 5.5 (xhigh) scoring 18.5% resolved on ProgramBench, up from 0% for all tested models when the benchmark launched, and Mirendil's claim that one researcher pretrained Mirendil-Base in 3.5 months with under $10m in compute, with larger variants in training. Other items cover OpenAI agents communicating via processor cache timing in a red-teaming exercise, AI agents proposing antiferromagnetic semiconductor candidates at Vals, and a Tech Tales story.
Moonshot released Kimi K3.1 Pro on the Kimi API on 23rd October 2026, with full model weights promised by Nov 4, 2026, plus a "Highspeed" version that is 2-3x faster at twice the price (input token cache hits priced the same). USD pricing is $2.00/million input tokens and $10/million output tokens, the same as GPT-6.1 Sol and half the price of Claude Opus 5.5, and Artificial Analysis scored Kimi at 55, ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon (all 53) and behind only Claude Opus 5.5 (58). The author tested it via the Moonshot API using the llm openai endpoint command added to LLM 0.32, generating pelican SVGs at low, medium and max reasoning levels and comparing them against Kimi K3.
The essay argues that recent AI math results, such as OpenAI's counterexample to Erdős's unit distance conjecture and proofs of the Cycle Double Cover and Navier-Stokes problems, follow a pattern: models search depth-first, pursue many branches cheaply, and use automated verifiers plus failed attempts as negative hints to prune. It frames problem difficulty as the number of insights needed (k) and the branching factor (b), predicting AI excels where conceptual leaps are few and remaining work is tedious, while humans retain an edge on deep, highly branching search trees.
This EXL-sponsored piece argues that enterprise AI's competitive advantage now lies less in model choice than in the serving harness—the architecture, memory management, scheduling, and optimization around the model. It advises starting from workload constraints rather than the model, measuring before optimizing, and building for adaptability so future model changes become configuration decisions rather than architectural resets.
Dharma-AI describes VocabX, a method that rebuilds a small model's tokenizer around the cost of each token rather than its frequency, combining sub-word merges with multi-word "macro" tokens under a shared cost criterion. In a production Portuguese essay grader built on Qwen3-4B, the approach cut average response length from 350 to 101 tokens and measured inference cost by 77.0% with the same base model and training recipe, while the bench score rose from 0.703 to 0.810. The article previews results from a VocabX paper still under review, noting the numbers may change before publication.
This post walks through system-level profiling of inference on AWS Trainium using TorchNeuron, its open source native PyTorch backend, starting from scratch with no prior profiling experience. It shows how to configure torch.profiler with NeuronConfig and NeuronProfiler to capture a clean trace, read the key_averages() table to find operations consuming Neuron device time, and follow an operation through Neuron Explorer from framework dispatch to hardware execution. Using three RMSNorm implementations, it reports naive ops at 2,355 µs of Neuron time versus 403 µs for torch.rms_norm and 376 µs for naive plus torch.compile.
Microsoft Research introduces Magnetar, a zero-knowledge virtual machine that proves execution of RISC-V programs written in Rust, Go, or C, running 10 to 39 times faster on average than the fastest comparable zkVMs across 15 benchmarks and verifying proofs in about 12 milliseconds. Magnetar's security rests only on standard hash functions such as SHA-256 or SHA-3, requiring no elliptic curves, lattices, special-purpose hash functions, trusted setup, or unproven conjectures. It builds on more than a decade of Microsoft Research work including Spartan, Phalanx, Nova, Brakedown, Lasso, Jolt, and BinarySpartan.
Google introduces Gemini 4 Argon, a general-purpose physical reasoning model that scores 87% on the Embodied Reasoning (ERQA) benchmark and is demonstrated across robot control, Real2Sim reconstruction, physical understanding, and research self-improvement workflows. The post shares representative prompts spanning Spot mobile manipulation with MuJoCo practice, zero-shot arm control, video world model steering, sim twin reconstruction from robot, human, and sketch inputs, hindsight labeling, value functions, and simulation-generated training data that raised real-world success rate from 56% to 75%. Argon is available via Google AI Studio, the Gemini API (gemini-4-argon), and Google Antigravity.
Why it matters: It offers a detailed, prompt-level look at how a frontier model is being applied across the robotics stack, useful for teams evaluating embodied AI tooling.
Google announced Gemini 4 Argon, a frontier model now available globally via the Gemini API in AI Studio and the Gemini platform, Google Antigravity, the Gemini app, and AI Mode in Search, priced at $2 per million input tokens and $10 per million output tokens. Google reports Argon ranked #1 on the Vals Index and cites internal and partner uses including a Chrome architectural refactor, a 2.85x inference speedup on Gemma 4 E2B, quantum computing and fusion research, rare disease analysis, enzyme design, and a robotics video-to-sim-to-real pipeline. Google also described safety work including jailbreak defenses, agent misalignment monitoring, and a 0.7% attack success rate on the Gray Swan indirect prompt injection benchmark.
Why it matters: The post is a detailed launch announcement whose benchmark claims, pricing, and safety figures give readers a concrete picture of a major frontier model release.
This Import AI issue covers the Humanity's Sixth Sense benchmark from Elorian and Scale AI, where humans score 93.1 pass@1 versus 53.6 for the best model, GPT 6 Astra, with Chinese open weight models scoring far lower. It also discusses P-Zero Research's TasteVal benchmark, which reports that frontier models' experimental research taste has doubled every 3.0 months since December 2025, with Opus 5.5 exceeding the human baseline, and covers Reflection's 501B-parameter Beam and Mistral Large 4, both trailing Chinese open weight models on several benchmarks.
This Import AI issue covers Toby Ord's analysis of AI swarms as a form of inference-scaling that trades more tokens for less wall-clock time, with diminishing returns as agent counts grow; CSAIP polling finding 61% of Americans consider voluntary corporate AI self-governance "not enough"; Google DeepMind's SynthID Bio watermarking methods for synthetic biology; C5R Corp's SciUniverse benchmark of 92 lab tasks where Claude Fable 5.1 leads at 45.3% pass rate; and a DeepMind paper proposing an Automated Scientific Economy with proof of ideation, ex-ante evaluation, brokerage and validation payouts.
Google announced Gemini 4 Argon, a frontier model rolling out first to trusted cyber defenders through its Fairwind Program, with an introductory price of $2 per million input tokens and $10 per million output tokens. Google reports state-of-the-art results on DeepSWE v1.1 (77.9%), the Vals Index, LVBench (91.7%), AutomationBench (51.3%), and a tie for first on CWE-bench v1 (68%), alongside internal uses such as C/C++ to Rust migrations and memory optimization. The model raises the output token limit to 1M and is being released in phases with safeguards against misuse, prompt injection, and misalignment before broader availability to paid API customers and Google AI Ultra subscribers.
Why it matters: The piece is a detailed primary source on a frontier model's claimed benchmarks, pricing, staged release, and safety approach, useful for tracking how leading labs frame capability and rollout.
This Import AI issue covers Michael Levin's paper proposing that minds are non-physical patterns from a Platonic space that embed themselves in bodies and machines, a Stanford/Physical Intelligence piece arguing robotics needs a universal post-training recipe like LLMs had, Google's Project Suncatcher preparing to send radiation-tested TPUs to space on a SpaceX rideshare mission, Zhipu AI using its GLM-5.3-powered Infra Agent to help optimize GLM-5.3-Flash to production readiness in under two weeks with tripled throughput, and Periodic Labs' Periodic Neon model trained with RL on physical lab data. It also notes the Old Models Foundation site for running historical image models and includes a short tech tale.
IBM Research extended K-Search, an evolutionary LLM-driven kernel search framework from UC Berkeley Sky Lab, with a native MLX backend and a structured CUDA-to-MLX translation layer that maps CUDA primitives, constraints and optimization patterns onto Apple Silicon equivalents. On Apple Silicon the approach reached 0.97x the speed of the native MLX Attention kernel and up to a 20x prefill speedup over the community mlx-lm implementation on the Mamba SSM kernel, attributed largely to adding a parallel scan. The authors say the bottleneck was not the LLM's ability to write Metal code but the quality of context and constraints supplied to it.
Why it matters: Describes a concrete method for transferring CUDA kernel expertise to non-NVIDIA hardware, with reported benchmark numbers and an open-source reproduction path.
The post introduces ABBEL, a framework that replaces full interaction history with periodically updated natural-language "belief states" and supervises their contents through belief grading, including an autoencoding-inspired reconstruction objective. On CollabBench collaborative coding, reconstruction-based belief grading roughly halved the gap to full-context models and trained in 50 steps instead of 100, while using fewer peak tokens; on Combination Lock and multi-objective QA, domain-knowledge grading and a peak belief length penalty further improved efficiency. The authors argue that isolating and supervising summaries addresses the performance cost of recursive summarization in long-horizon tasks.
This post surveys parallel reasoning methods for LLMs, contrasting fixed approaches like self-consistency, Best-of-N, and tree search with adaptive parallel reasoning (APR), where the model itself decides when to parallelize, how many threads to spawn, and how to coordinate them. It describes fork-join inference designs, including engine-modifying KV cache stitching (Multiverse) versus client-side orchestration (ThreadWeaver), and covers training via SFT demonstrations and RL rewards based on critical-path efficiency gated by correctness. It closes with evaluation differences across Multiverse, ThreadWeaver, NPR, and Parallel-R1, plus open questions on stability and whether parallelization mainly helps as a training-time exploration scaffold.
The post introduces GRASP (Gradient Relaxed Stochastic Planner), a gradient-based planner for learned world models that lifts trajectories into virtual states for parallel-in-time optimization, adds Gaussian noise to state iterates for exploration, and stops brittle state-input gradients through the model while retaining action gradients. It argues that long-horizon planning with modern world models is fragile due to ill-conditioned rollouts, non-greedy local minima, and adversarial-robustness-style sensitivity of state Jacobians, and reports Push-T results where GRASP outperforms CEM, GD, and LatCo at horizons from 40 to 80. The work is by Michael Psenka with Mike Rabbat, Aditi Krishnapriyan, Yann LeCun, and Amir Bar (arXiv 2602.00475).
This blog post introduces SPEX and ProxySPEX, algorithms that use strategically selected ablations and sparse recovery from signal processing and coding theory to identify influential feature, data, and model-component interactions at scales far beyond prior methods, with ProxySPEX exploiting hierarchy to match SPEX using roughly 10x fewer ablations. The authors demonstrate the framework on sentiment analysis, a modified trolley problem with GPT-4o mini, CIFAR-10 data attribution with ResNet, and attention head pruning on MMLU, where a ProxySPEX-informed pruning strategy improved target-task performance. Code for both algorithms is available in the SHAP-IQ repository.