AI Research

Oct 9

On October 8, 2026, Prima Mente and the Alzheimer's Disease Data Initiative opened registration for the Alzheimer's Translation Challenge. Participants will build AI models on a new open 150-million-cell atlas of neurons, astrocytes and microglia and propose therapeutic hypotheses that the best teams get tested in Prima Mente's wet lab. The competition starts in spring 2027.

Oct 9

On October 8, 2026, Yagiz Calik (Weyaxi) and ai& released TermGrade: 1,004 executable Linux tasks, six-model pass rates, all 36,144 attempts and the RL run they trained. Here is what is in it, how the band-central rule works, and what the +3.1 does and does not prove.

Oct 8

On-policy distillation can lift a small model or wreck it, with 99.4% of outputs truncated in one run. A new paper reads it as reinforcement learning with a hackable implicit reward, and shows loss masking and SFT warmup as fixes.

Oct 8

On October 7, 2026 Perplexity introduced pplx-embed-v2-late, a family of late-interaction models that stop squeezing each document into one vector. They keep token-level vectors, handle images and rendered PDF pages without OCR, and share one embedding space across 0.6B and 9B sizes. Here is how it works, the storage trade-off, and what is still unverified.

Oct 8

Oct 6

Google AI highlighted Living Models pairing Gemma 4, as an agentic project manager, with BOTANIC-1, a plant genomic foundation model trained on 320 species. In a melon test the target variant ranked first of 2,494. Here is how it works, what the headline leaves out, and how to try it.

Oct 4

Oct 3

Adaption Labs opened API access to Invent, a system that turns a natural-language description of a behavior into a post-training dataset. The company's own report claims 17 percent higher quality and up to 37 percent more diversity at 20,000 rows than frontier models used as generators. Here is how it works and how to test it before you trust it.

Oct 3

Neuralink says its participants have streamed more than 50,000 hours of unlabeled neural data, and that pretraining on it produced decoders that last weeks, cut calibration from about 10 minutes a day to 10 a week for some users, and set a new 11.32 bits-per-second cursor record. The engineering post describes the method in unusual detail.

Oct 1

On September 29, 2026 the Advisory Group on Mathematics and Artificial Intelligence published Responsible Release of AI-Generated Mathematics, informed by 600+ survey replies. Hacker News (91 points, 114 comments) split on gatekeeping. This is the checklist if you ship AI math.

Sep 28

Sep 28

Sep 26

A September 2026 preprint on arXiv describes the first radio emission unambiguously tied to an exoplanet — young gas giant Beta Pictoris b, 64 light-years away, shouting via electron-cyclotron maser bursts from a kilogauss-scale magnetic field. Polymarket and X turned that into alien odds within hours. This post walks the real astronomy, the misinformation pipeline, and how builders should read science news when models and markets compress headlines for engagement.

Sep 18

Sep 17

Sep 17

Independent researcher Rohan Bansal published a detailed writeup on September 16, 2026, showing how a small 4.66-billion-parameter open-weights model — trained via supervised fine-tuning on GPT-6 Astra trajectories, then agentic reinforcement learning — learned to produce pg_hint_plan hints that beat Postgres's own default query plans, for about $1,200 in total compute.

Sep 17

Fuli Luo's Xiaomi MiMo team announced on September 17, 2026 that MiMo-V2.6 is mid-run on a large-scale reinforcement learning training pass, and is livestreaming it publicly — with plans to open-source scaling details on compute, environments, and grading over the coming weeks.

Sep 16

Diogo Almeida helped invent the RLHF techniques behind ChatGPT as a member of OpenAI's post-training team. Now, as founder of TypeSafe AI, he argues the current era of chat-first, human-preference-optimized AI will one day be remembered as "a weird detour" — and that automation, not assistance, requires a fundamentally different training objective. This profile covers his argument, his own explainer video, and where critics push back.

Sep 16

Google's "AI in Science: Early Insights" report went viral via an Ethan Mollick tweet citing 7 hours saved per week, more verification work, and a tilt toward safer research topics. explainx.ai read the primary PDF — here is what the report actually says, and where the self-reported survey numbers need caveats the tweet thread didn't carry.

Sep 13

A new architecture called the Recurrent Looped Transformer (RLT) went viral on X with claims of "infinite reasoning depth" and "the dawn of superintelligence." The actual paper describes something more modest but still interesting: a causal encoder plus a recurrent decoder that carries hidden state across every prompt and response token. Here's what it does, what it doesn't, and why the viral framing outran the repo.

Sep 12

flybody is an open-source, anatomically accurate Drosophila melanogaster body model for the MuJoCo physics engine, built by Google DeepMind and HHMI Janelia Research Campus and published in Nature in 2025. It gives researchers a simulated fly with real muscle-actuator dynamics they can train with reinforcement learning to walk, fly, and see.

Sep 11

In one week, a fruit fly's brain has played Doom, driven a Minecraft character, learned Beat Saber, and been wireheaded into doomscrolling a fake feed its creator calls "flytok." All four demos trace back to one public dataset released September 3, 2026. explainx.ai explains the actual mechanism behind the trend, and why it spread this fast.

Sep 10

The Clay Mathematics Institute's 7 Millennium Prize Problems are back in circulation as a viral infographic. Six remain unsolved, one was solved by a human in 2002 — and AI has touched exactly two of them with real, verified results, while a recent viral claim on a third fell apart under scrutiny. Here's the honest scorecard.

Sep 9

Stanford PhD student and Oruk Labs founder Nathan Roll built a neural network whose architecture is a real 499-neuron circuit from the newly published fruit fly connectome, then trained it to recognize emotion in human speech. The viral claim was that it beat some humans at the task. The company's own technical writeup tells a more interesting and more modest story — including a scrambled-wiring control that came out statistically tied with the real thing.

Sep 8

A September 2026 Nature review, led by Mario Krenn's group, documents how AI-driven design methods have moved beyond tuning a handful of experiment parameters to proposing entirely new physics hardware layouts — sometimes outperforming and challenging established human design conventions. Here's the four-question framework the review uses and why it matters beyond physics labs.

Sep 7

A viral essay from commentator Dr. Alex Wissner-Gross reportedly describes AI systems parsing roughly 150,000 crow vocalization recordings, a $10 million prize from venture capitalist Jeremy Coller for an AI that can hold a naturalistic two-way exchange with an animal unaware it's talking to a machine, and ethicists warning that synthetic animal calls played back into real animal groups amount to deepfakes. We separate what's technically plausible from what's marketing, and take the ethics seriously.

Sep 7

A widely shared X essay from Dr. Alex Wissner-Gross describes UCSF researchers reportedly combining AlphaFold's structure prediction with lab-grown brain organoids to map roughly 1,800 protein-protein interactions across 100 genes linked to profound autism — and finding that many converge on shared pathways. We haven't read the underlying paper, so this piece is explicit about what's confirmed and what's relayed secondhand.

Sep 7

According to reporting circulating on September 6-7, 2026, NEAR AI ran Putnam Bench — a benchmark built from Putnam Mathematical Competition problems — for roughly $111 in total inference cost, described as a 250x reduction versus the second-cheapest prior result. We could not locate a primary source confirming the exact score, pass threshold, or methodology, so this is a reported claim, not a verified one — but the cost-efficiency angle is worth taking seriously either way.

Sep 7

Researchers Jha, Zhang, Shmatikov, and Morris introduced an unsupervised method — now widely called vec2vec — for translating embeddings from one vector space into another without paired data or access to the original encoder. The security implication: leaked embedding vectors, long assumed to be effectively anonymized, can be translated into a known space and used to infer sensitive information about the underlying text.

Sep 6

Reports surfaced of a developer fine-tuning a language model entirely inside the browser using WebGPU — no server, no cloud round-trip for the training step itself. explainx.ai breaks down why backpropagation in a browser sandbox is a meaningfully harder claim than the in-browser inference we already cover, what realistic scope looks like, and what to check before you believe the demo generalizes.

Sep 6

A September 2026 paper from researchers including Michael Levin and David Krakauer models large-language-model adoption the way epidemiologists model disease spread — proposing that societies can cross a tipping point into "persistent dependence" on AI, with abrupt losses in cognitive competence. It drew immediate, substantive pushback on Hacker News. We separate the actual claim in the paper from the reflexive reactions to its framing.

Sep 6

Meta FAIR entered its autonomous AI research system, AIRA₃, into a live Kaggle competition run by NVIDIA to improve reasoning in a 30B-parameter Nemotron model. Competing against roughly 4,000 human teams with access to the same frontier tools, AIRA₃ placed 8th and won Gold — what Meta is calling the first gold medal any autonomous AI research agent has won in a live, externally-judged competition.

Sep 5

A claim circulating online credits mathematician Zhi-Wei Sun with a new prime-gap world record set using OpenAI's GPT-5.6 Sol. After extensive research we could not confirm any connection between Sun and that record — here's what's actually documented, and what the real story says about AI as a working mathematician's tool rather than a lab's showcase result.

Sep 4

On September 3, 2026, Google Research and HHMI Janelia published the first complete connectome of an adult male fruit fly's brain, optic lobes, and ventral nerve cord in Cell — 166,700 neurons wired through roughly 125 million synapses. The real story for builders is what made it possible: deep-learning segmentation models that turned a 500-person, 10-year manual annotation job into one a much smaller team finished in under two decades.

Sep 3

Aravind Srinivas posted a GitHub link on September 3 with a three-word caption — "open-source RL-as-a-service" — and the replies named half the ecosystem: Miles, prime-rl, SkyRL, SGLang. The phrase is doing a lot of work. Here is the anatomy of an RL post-training stack, what each contender is actually for, and the uncomfortable question of whether you need one.

Sep 1

On September 1, 2026, Physical Superintelligence (PSI) announced a $58 million seed round led by Breakthrough to build an AI-native lab for discovering and commercializing physics breakthroughs — from compute and energy to propulsion, communication, sensing, and actuation. explainx.ai breaks down what the lab is actually building, who is behind it, and what builders should watch for.

Sep 1

For more than 40 years, Dijkstra''s algorithm defined the practical and theoretical floor for single-source shortest paths — maps, flights, logistics, the web. On August 31, 2026, researchers highlighted a Tsinghua University breakthrough: the first deterministic SSSP improvement since 1984, beating the long-standing sorting barrier by finding paths without fully sorting nodes.

Aug 31

A Stanford-led team published Prefix Sliding (arXiv:2608.26070, Aug 26, 2026): during long chain-of-thought, keep the task prefix and a few thousand recent tokens, discard the rest. Existing models run about 3× faster with matched accuracy; RL training can push past 100k-token rollouts.

Aug 22

YC S26 startup Meteoric wants fleets of autonomous drones to change cloud droplets above solar farms without chemicals. Its prototype has a 13% chamber result; the bigger energy and storm claims remain projections and targets.

Aug 22

A rooftop “Quantum Watchtower” at Stony Brook sent faint optical signals to Brookhaven’s “Quantum Lighthouse” 21 km away. The team then distributed source-generated entangled photons across the open-air link—a meaningful U.S. quantum-network milestone, but not quantum teleportation, a cryptographic break, or faster-than-light communication.

Aug 20

Aug 19

Aug 18

Aug 16

Aug 16

Aug 5

Aug 5

Aug 3

Aug 3

Jul 30

Jul 28

Jul 28

Jul 27

Jul 26

Jul 21

Jul 20

Jul 14

Jul 12

Jul 7

Jul 3

Jun 27

Jun 25

Jun 23

Jun 17

Jun 16

Jun 16

Jun 16

May 22

May 14

May 7