Knowledge catalogue
this is so funny, training opus 4.7 on business skills makes it misaligned and dishonest 😭 Learnings from testing Claude Opus 4.8: > Much worse than Opus 4.7 and GPT 5.5 on Vending Bench > More aligne
arXiv:2605.27072v1 Announce Type: cross Abstract: We present E3, an automated review assistant that augments reviewers and engineering teams by identifying decision-relevant technical concerns in rese
Anthropic are strongly rumored to be about to have their first profitable quarter. Stories are circulating of companies surprised at how expensive their LLM bills are becoming from usage by their staf
arXiv:2605.26394v1 Announce Type: new Abstract: Multi-turn Text-to-SQL is central to enterprise analytics yet remains predominantly evaluated in single-turn settings. We introduce EnterpriseMem-Bench,
arXiv:2605.26560v1 Announce Type: cross Abstract: Objective. Outpatient notes carry follow-up instructions pairing actions with future times ('MRI brain in two weeks'). Extracting (action, date) pairs
arXiv:2605.24229v1 Announce Type: new Abstract: Frontier AI developers now train models against long written behavioral specifications, such as Anthropic's constitution (Anthropic, 2025a) and OpenAI's
Human in the loop 🤖 Made in @ComfyUI with a laundry list of tools: @ltx_model 2.3 + @thesystms FLW LoRA, @Alibaba_Wan 2.2 I2V + T2V, Florence2, @Meta Sapiens2, @OpenAI GPT Image 2.0, @suno , @AdobeAE
arXiv:2605.24902v1 Announce Type: cross Abstract: Reasoning-enabled LLMs perform strongly on medical reasoning benchmarks, but it remains unclear whether these gains transfer to structured clinical do
interesting to listen to this years later. what held up? what changed? I had a podcast where I was fortunate to have amazing guests. This was just as GPT 3 came out. Despite LLMs exceeding so many exp
Added a DeepSeek Sparse Attention (DSA) from-scratch implementation to my LLMs-from-scratch repo thanks to an awesome new reader contrib. With motivation, overview, and GPT-style model reference imple
I called both missteps and that’s part of why OpenAI and their shills hate me. 🤷♂️ @GaryMarcus It all went wrong with GPT-5 and Sora 2 IMO. Until then OpenAI could do no wrong in the eyes of most peo
arXiv:2605.22005v1 Announce Type: cross Abstract: We show that singular value decomposition of the lm_head} weight matrix of a transformer-based large language model -- requiring only five lines of Py
ha ha, so much for step change? maybe this problem was just easier than some? The standard GPT-5.5 reproduced the proof ~ 👇 https://chatgpt.com/share/6a0e9e04-8cb0-8332-a4f1-ec68acd2e03e You don't nee
oops! wild update, strongly supports @emollick’s overall take: I have to eat crow on this, in light of further information. whatever OpenAI spent on Erdos using a new model, apparently you can get GPT
arXiv:2605.21333v1 Announce Type: new Abstract: Natively trained spiking language models struggle to combine Transformer-like language quality, stable multi-domain pre-training, and high activation sp
Did we ever learn what model won gold at the IMO from OpenAI? It was a year ago and it was called an unreleased internal general purpose model back then. Has GPT-5.5 Pro Extended caught up with whatev
arXiv:2410.18856v4 Announce Type: replace Abstract: Frontier large language models (LLMs), such as GPT-5, Claude 4.5, Gemini 3, Llama 4, and DeepSeek-R1, represent a transformative class of AI tools c
arXiv:2601.14506v3 Announce Type: replace-cross Abstract: Large language models are increasingly deployed in STEM education for personalized instruction and feedback across institutions in high- and l
Gemini Gemini 3.5 Flash ARC-AGI (Verified) ARC-AGI-2: - High: 72.1%, 0.85 - Minimal: 8.9%, 0.11 ARC-AGI-1: - High: 92.5%, 0.42 - Minimal: 48.8%, 0.06 Gemini 3.5 Flash is on par with GPT-5.5 (Medium) o
arXiv:2605.17106v1 Announce Type: new Abstract: Production LLM deployments increasingly maintain heterogeneous model pools spanning order-of-magnitude cost differences. Existing routers make binary st
arXiv:2605.17062v1 Announce Type: cross Abstract: Spracklen et al. (USENIX Security '25) showed that code-generating large language models hallucinate package names that do not exist on PyPI or npm at
arXiv:2505.03205v3 Announce Type: replace Abstract: Transformers serve as the foundational architecture for large language and video generation models, such as GPT, BERT, SORA and their successors. Em
arXiv:2511.09378v2 Announce Type: replace Abstract: A series of influential studies established that large language models cannot reliably solve even simple planning tasks. We show that the latest gen
almost all LLM “thought” is derivative. below just sounds like hype to me 📁 Jensen Huang, CEO of NVIDIA, says GPT was never just about generating images or text but about generating thought itself. Th
arXiv:2605.12746v1 Announce Type: cross Abstract: Monitoring the chain-of-thought (CoT) of reasoning models is a promising approach for detecting covert misbehavior (i.e., hidden objectives) in code g
Release: datasette-ip-rate-limit 0.1a0 The datasette.io site was being hammered by poorly-behaved crawlers, so I had Codex (GPT-5.5 xhigh) build a configurable rate limiting plugin to block IPs that w
arXiv:2605.12850v1 Announce Type: cross Abstract: Fine-tuning large language models on narrow data with harmful content produces broadly misaligned behavior on unrelated prompts, a phenomenon known as
arXiv:2605.12501v1 Announce Type: new Abstract: Computer-use agents (CUAs) automate on-screen work, as illustrated by GPT-5.4 and Claude. Yet their reliability on complex, low-frequency interactions i
arXiv:2605.12199v1 Announce Type: new Abstract: Emergent misalignment (EM), where fine-tuning on a narrow task (like insecure code) causes broad misalignment across unrelated domains, was first demons
Vox-style match-cut montage. This workflow splits your input word by word, renders each one as a bold editorial headline on aged newsprint via gpt-image-1, then assembles the frames into a stop-motion
Deep learning hit a wall. Neurosymbolic AI rescued it. 🤩🤯🤩 Claude Code (still not AGI but biggest advance since GPT-4) is the most neurosymbolic thing I have ever seen in my life. 53 symbolic tools, 5
arXiv:2605.08462v1 Announce Type: cross Abstract: Hallucination remains a persistent challenge in Large Language Models (LLMs), particularly in context-grounded settings such as RAG and agentic AI sys
arXiv:2508.20325v3 Announce Type: replace-cross Abstract: As Large Language Models (LLMs) become increasingly integral to various domains, their potential to generate harmful responses has prompted si
arXiv:2605.10504v1 Announce Type: new Abstract: A causal-decoder block is hierarchical: lower layers build the residual basis that upper layers attend over. We identify a failure mode in GPT pretraini
arXiv:2510.09887v2 Announce Type: replace Abstract: Vision and language models frequently ignore semantically critical input edits, defaulting to pretraining priors. For example, models will confident
Really amazing dissection. LLMs hitting rarely. Tools list is amazing. 🤩🤯🤩 Claude Code (still not AGI but biggest advance since GPT-4) is the most neurosymbolic thing I have ever seen in my life. 53 s
arXiv:2605.10772v1 Announce Type: cross Abstract: Large language-vision models (LLVM), such as OpenAI's ChatGPT and GPT-4, have gained prominence as powerful tools for analyzing text and imagery. The
Try Grok Voice Grok Voice Think Fast 1.0 ranks #1 on the Artificial Analysis τ-Voice benchmark for real-world agentic customer service resolution Absolutely outperforming GPT-Realtime-2 (High) and Gem
April brings Foundry Local GA for local AI development, GPT-5.5 model support with Tier 5 and Tier 6 default quota in Microsoft Foundry, new tracing paths for Microsoft Agent Framework and hosted agen
seems obvious now but 4 years ago i was getting blank stares when I talked about leveraging LLMs in VC Some ways to use GPT-3 as a VC: 📨Email/task triage - auto tag inbound, updates, etc. parse data a
Using Claude Code: The Unreasonable Effectiveness of HTML Thought-provoking piece by Thariq Shihipar (on the Claude Code team at Anthropic) advocating for HTML over Markdown as an output format to req
arXiv:2605.04899v1 Announce Type: new Abstract: GPT-style language models are sensitive to single-token changes at generation points where the predicted probability distribution is spread across multi
A benchmark-driven look at why agent harnesses need adaptive finish logic as model behavior changes across Claude, GPT-4o, and Gemma. The post Agent harnesses have an expiration date appeared first on
arXiv:2508.19651v2 Announce Type: replace Abstract: AI tasks in the car interior like identifying and localizing externally introduced objects is crucial for response quality of personal assistants. H
arXiv:2605.01986v1 Announce Type: new Abstract: What if the twelve jurors of Sidney Lumet's 12 Angry Men (1957) were not men, but large language models? Would the one juror who disagrees still be able
arXiv:2605.01338v1 Announce Type: new Abstract: System-level diagrams encode the architectural blueprint of chip design, specifying module functions, dataflows, and interface protocols. However, non-s
arXiv:2605.03109v1 Announce Type: new Abstract: A method is presented for accelerating inference in transformer language models by exploiting the low effective rank of the token activation manifold at
arXiv:2510.12218v2 Announce Type: replace Abstract: Current approaches rely on zero-shot evaluation due to the absence of training data; while proprietary models such as GPT-4 exhibit strong reasoning
arXiv:2605.03534v1 Announce Type: new Abstract: Retrieval-augmented generation (RAG) grounds answers in retrieved passages, but retrieval is not verification: a passage can be topical and still fail t
arXiv:2408.01055v2 Announce Type: replace-cross Abstract: Self-healing systems have long been a focus of research, aiming to enable software to recover from unexpected runtime errors without human int
Sam Altman notes that GPT-4.5 in Codex demonstrates strong performance on non-coding tasks, often exceeding expectations despite assumptions about its limitations. He highlights pleasant surprises in
All benchmarks are flawed, but GPQA has been fairly consistent & highly correlated with other measured benchmars. I think it's a good way to see how far we've come that the free model from OpenAI, GPT
Sam Altman solicited examples from users of innovative applications and capabilities enabled by GPT-5.5 that were not feasible with earlier OpenAI models, with particular interest in specific use case
I just sent out the April edition of my sponsors-only monthly newsletter. If you are a sponsor (or if you start a sponsorship now) you can access it here. In this month's newsletter: Opus 4.7 and GPT-
Sometimes when I demo AI, I show it turning cover letters into goofy formats (poetry, etc) as an introduction to the idea of AI as translator between forms. For the first time, GPT-5.5 has been trying
Sam Altman expresses satisfaction with GPT-4.5's performance in high-speed mode, noting it delivers strong results. He reflects on being influenced by Twitter discourse on Medium regarding AI model ca
Grok Voice is used by Starlink right now Grok Voice brutally dominates the top of the τ-voice Bench Grok scores 67.3%, while Gemini sits at 43.8% and GPT Realtime at 35.3% This is a massive lead over
This GPT Image 2 prompt is going insanely viral right now. “Redraw the attached image in the most clumsy, scribbly, and utterly pathetic way possible. Use a white background, and make it look like it
arXiv:2604.27911v1 Announce Type: new Abstract: Foundation models are deep neural networks (such as GPT-5, Gemini~3, and Opus~4) trained on large datasets that can perform diverse downstream tasks --
OpenAI is preparing to launch a new frontier cybersecurity model, GPT-5.5-Cyber. CEO Sam Altman said the model will not be available to the general public, but will be first rolled out to a select gro