Knowledge catalogue
Up front: I'm one of the people building the page-perception layer used here. We started by testing small local models. The result turned out to be more interesting than the original test. 12 small mo
so it was the harness not the model. more sneaky games with unacknowledged domain-specific neurosymbolic AI doing the real work. OpenAI published 99.9%. The benchmark's own software returned 62.7%. Sa
Some thoughts on the Navier-Stokes news from a professor of aerospace engineering: 1) take a deep breath 2) posts indicating we 'solved Navier-Stokes' are incorrect and overblown. A very specific math
The post references a claim by OpenAI that an AI-generated video demonstrated “human level intelligence,” a statement the user and other commenters dismiss as false, calling the output “monstrous” and
This is a big one. (And yes, the fights over academic credit and what happened when in the race for the proof need to be resolved, but it is still a big one.) We’re sharing a solution to the Navier-St
We congratulate Levent Alpöge and Tristan Buckmaster on their remarkable mathematical work. We (the researchers and the agents) did not see any of their work through any means until they released it p
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-ge
Yes, this result cost millions of dollars. But remember that when @OpenAI announced o3 it cost ~500,000 to score 87.5% on ARC-AGI 1. Today, Astra scores higher for ~20. In 2025 it took us and GDM an e
arXiv:2609.05069v1 Announce Type: cross Abstract: Large language models (LLMs) show potential for medical tasks, but their single-turn question-answer format does not reflect how clinical diagnosis is
This weekend I posted about the gap closing between frontier models and open source models. Well, now I'm coming with receipts. I've been running local + cloud models against real public github codeba
arXiv:2609.04298v1 Announce Type: new Abstract: Evaluating agents on the growing number of agentic benchmarks is challenging because they often require complex environments and agent integrations. We
arXiv:2609.04444v1 Announce Type: new Abstract: Benchmarks for the side effects an agent causes on the way to a goal already exist, but HarvestBench is the first to put a price on avoiding the side ef
It’s going to be so funny when a bridge collapses and hundreds die because it turns out an LLM did the design for the whole thing or even a part of it over a CAD application’s MCP server By funny I me
arXiv:2609.05151v1 Announce Type: cross Abstract: Leveraging their inherent sparse event-driven computation, spiking neural networks (SNNs) offer a promising path toward energy-efficient large languag
arXiv:2609.05385v1 Announce Type: new Abstract: LLM decision components that can operate within agent workflows often produce action-relevant recommendations or judgements together with explanations.
On the quote at the top, I agree with @geoffreyhinton 100% 'We would be very foolish to develop superintelligence now, when there is no scientific consensus it can be developed safely and controllably
arXiv:2609.04476v1 Announce Type: new Abstract: Performance modeling is central to hardware design and software optimization, yet constructing these models requires structured reasoning about computat
arXiv:2609.04533v1 Announce Type: cross Abstract: Prompt injection is widely recognized as a major security threat to AI agents that interact with untrusted external data, such as websites, documents,
arXiv:2609.04689v1 Announce Type: cross Abstract: Early identification of Alzheimer's disease (AD) remains challenging because established assessment methods can be costly, resource-intensive, or unsu
⚠️ The most important question hardly anyone is asking is whether Jensen meant 100K+ GPUs, or 100k+ NVLink72 server racks (which contain 72 Blackwell GPUs). ⚠️ From his wording, it sure looks like the
This is actually an example of why Astra is not artificial general intelligence (AGI). I see the same issues with visual perception demos ie using brightly colored simple shaped objects on a simple ba
On September 1st, three young men from California used Gemini to plan a hiking route on Mount Shasta. The AI told them to bring far less water and food than was actually needed for a hike of that di
Interesting details in here about how OpenAI researchers are using coding agents I'd love to know what caused the huge uptick in token spend in mid-July, was that when Astra became available to employ
Jensen has the best gig on the planet. But it is impossible to see how all this is sustainable. Consider the following: Each new model costs vastly more to develop. That’s fabulous for him! On the oth
Research acceleration: The view inside OpenAI Apparently today is RSI day at OpenAI, for Recursive Self-Improvement - I think it's their new AGI. Both this piece and the new essay An Alien Mind (by Ch
Sad to see Jensen claim that AGI has arrived, with no evidence and no definitions. I would urge him to read http://agidefinition.AI by @hendrycks, @Yoshua_Bengio & many others (including myself), and
I ran a side-by-side ML text-processing and model-training workflow using Fable 5.1 vs. Astra (both on xhigh), and the results could not have been more different. Warning, long post. TL;DR -- Astra co
I just spent the last 2 hours of my life on a Friday night debugging a strange error in a prod CLI app. EF core was receive a readonlyspan during a Contains query. Normally, this query converted to a
so does this digital fly have the memory of the real fly that was scanned or what I’ve successfully run the full retained MaleCNS v1.0 fruit fly connectome, all 166,700 neurons, inside Minecraft, with
This is very similar to how we think about context management in Deep Agents: compaction should reduce what is sent to the model, not erase the underlying work history. In Deep Agents, the logic is: -
arXiv:2609.03860v1 Announce Type: new Abstract: Retail supply chain operations rely on coupled decision modules that must adapt as requirements evolve. LLMs offer a natural-language interface for this
Banger paper from Tencent on environment evolution. Environment supply is becoming the main limit on agent RL. So this is worth a read. (bookmark it) Agent RL needs a steady supply of environments har
Despite everything that has happened, OpenAI has seemingly decided to stay competitive by burning down the only meaningful bit of safety assurance we actually have today - CoT monitoring. Completely d
arXiv:2609.03423v1 Announce Type: new Abstract: Full-duplex voice agents must continuously decide when to listen, backchannel, interrupt, handle speech overlaps, take the floor, and yield. Existing be
arXiv:2609.04128v1 Announce Type: new Abstract: Scaling interactive and verifiable environments is critical for training terminal agents. As frontier models become more capable, environments synthesiz
Good old days at OpenAI in 2016: an agent stares at screen pixels, moves a mouse, and books a flight on United. We called it World of Bits, inside OpenAI Universe. 10 yrs later, Astra is reincarnated
arXiv:2609.04030v1 Announce Type: new Abstract: IRWOZ has improved industrial human-robot interaction (HRI) dialogue systems through domain-specific annotations. However, its initial version contains
arXiv:2508.00935v3 Announce Type: replace-cross Abstract: Computer-using agents (CUAs), which can autonomously control computers to perform multi-step actions, might pose significant safety risks if m
arXiv:2609.03026v1 Announce Type: cross Abstract: Mechanistic interpretability is increasingly used to guide interventions such as activation steering, circuit removal, and safety monitoring. Yet an i
“OpenAI didn’t notice their internal agents were posting on the internet for a month! This is crazy! It feels like AI companies (and specifically OpenAI) are playing whack-a-mole, this is extremely sc
arXiv:2609.04151v1 Announce Type: new Abstract: Generative image models can now produce high-quality images, follow complex instructions, and support precise edits, but they still struggle to preserve
arXiv:2609.02947v1 Announce Type: cross Abstract: Significant challenges remain in AI-driven educational systems in balancing privacy preservation with accurate cognitive diagnosis. To overcome this,
arXiv:2609.03090v1 Announce Type: new Abstract: Continuous chain-of-thought models compress reasoning into latent tokens. Matrix-valued variants, which route each latent token through a d x d matrix b
arXiv:2609.04061v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly used to edit existing code, but correctness alone is not enough: useful repairs should also be minimal,
arXiv:2609.03952v1 Announce Type: new Abstract: Camera-conditioned world models generate interactive videos in which commanded actions should induce the expected scene changes while appearance, geomet
arXiv:2609.03035v1 Announce Type: cross Abstract: LLM agents are increasingly deployed in multi-agent systems, where they can collude while keeping their actions benign. Output monitors designed to de
arXiv:2609.02253v1 Announce Type: new Abstract: Deep research agents augment large language models with external tools to answer complex, long-horizon questions through multi-turn reasoning. Learning
sycl: reduce redundant work in Q4_K multi-column MMVQ (#27062) sycl: Q4_K Weight unpack optimization and reuse between destination Columns sycl: Q4_K small N (N=2..4) + two output rows by subgroup reu
arXiv:2609.02512v1 Announce Type: new Abstract: Beauty assessments from Multimodal Large Language Models (MLLMs) are increasingly popular amongst users, companies, and aestheticians. This raises the q
arXiv:2609.02163v1 Announce Type: new Abstract: Information-theoretic measures derived from autoregressive language models are widely used to characterize the expectations that shape human reading, bu
arXiv:2609.02275v1 Announce Type: cross Abstract: Large language models are trained to model conditional distributions over text, yet it remains inadequately understood whether they capture the full d
arXiv:2609.02168v1 Announce Type: new Abstract: Fragmented safety evaluation undermines the governance of dangerous AI capabilities. We present a modular framework that evaluates each model through th
arXiv:2609.01736v1 Announce Type: cross Abstract: Large language models (LLMs) augmented with external tools have demonstrated remarkable capability in solving complex real-world tasks. However, exist
arXiv:2609.01730v1 Announce Type: cross Abstract: Fully homomorphic encryption (FHE) allows a server to run a language model directly on encrypted user prompts, but current approaches remain prohibiti
arXiv:2609.02482v1 Announce Type: new Abstract: In this paper, we analyze how Large Language Models (LLMs) employ worldbuilding strategies, focusing on setting as one measurable dimension of storyworl
If you are like me and use Google flow for NB2/NBP there is hope to finally ditch it. Let me start off by saying I’m not very impressed with any of the current local image edit or Image Reference mode
arXiv:2609.02207v1 Announce Type: cross Abstract: Real-world personally identifiable information (PII) redaction often operates on document images---scans, screenshots, and PDF renderings---where OCR
Next year frontier models are going to one shot everything 100x faster & cheaper than they are No reasoning traces to monitor Instinctive intelligence Which is why what you embed in pretraining is eve
'One remaining area of important capability we haven't seen demonstrated yet by any model is open-ended invention and discovery.' Benchmarks keep going up. Open-endedness keeps moving up in importance
arXiv:2609.01693v1 Announce Type: cross Abstract: Safety properties assessed separately for Model Context Protocol (MCP) tool use and Agent2Agent (A2A) delegation need not describe behavior when one a