31
Claude Code daily benchmarks for degradation tracking
Track Claude Code's daily performance on SWE-Bench-Pro as we collect an Opus 5.5/high baseline.
Most cited
Posts that commenters keep linking to, ranked by how many different people have cited them in Hacker News comments. 1,013 posts.
32
Software 2.0 (2017)
I sometimes see people refer to neural networks as just “another tool in your machine learning toolbox”. They have some pros and cons, they…
33
The AI-Box Experiment
On the first occasion (in March 2002), Eliezer Yudkowsky simulated the AI and Nathan Russell simulated the gatekeeper. The AI’s handicap (the amount paid by the AI party to the gatekeeper party if not released) was set at $10. On the second occasion (in July 2002), Eliezer Yudkowsky simulated the AI and David…
34
DALL·E: Creating Images from Text
We’ve trained a neural network called DALL·E that creates images from text captions for a wide range of concepts expressible in natural language.
35
A postmortem of three recent issues
A technical report on three bugs that intermittently degraded Claude's responses: what happened, why fixes took time, and what we're changing.
36
Measuring AI Ability to Complete Long Tasks
We propose measuring AI performance in terms of the length of tasks AI agents can complete. We show that this metric has been consistently exponentially increasing over the past 6 years, with a doubling time of around 7 months. Extrapolating this trend predicts that, in under a decade, we will see AI agents that can…
37
How a Kalman filter works, in pictures
I have to tell you about the Kalman filter, because what it does is pretty damn amazing.
38
We built voice modulation to mask gender in technical interviews
We built software to mask gender in technical engineering interviews. See what we learned and why it matters.
39
Assessing Claude Mythos Preview's cybersecurity capabilities
Technical details on how we tested Claude Mythos Preview, a model strikingly capable at computer security tasks, and what we found.
40
Deploying Transformers on the Apple Neural Engine
An increasing number of the machine learning (ML) models we build at Apple each year are either partly or fully adopting the Transformer…
41
Claude Code: Best practices for agentic coding
Get more out of Claude Code: tell Claude what it can't see, give it a check to run, and enforce the rules that can't have exceptions.
42
Jukebox
We’re introducing Jukebox, a neural net that generates music, including rudimentary singing, as raw audio in a variety of genres and artist styles. We’re releasing the model weights and code, along with a tool to explore the generated samples.
43
Why I Am Not An Integrated Information Theorist (or, The Unconscious Expander)
Happy birthday to me! Recently, lots of people have been asking me what I think about IIT—no, not the Indian Institutes of Technology, but Integrated Information Theory, a widely-discussed &#…
44
AI Written, AI Read cartoon
One piece of slang that has long embodied the short attention span Internet age is TL;DR, short for “too long; didn’t read.” With the explosion of generative AI tools, we’re rapidly entering the age of TL;DW: “too long, didn’t write.” A January survey from Fishbowl found that 40% of nearly 12,000 workers have used…
45
AA-Omniscience v1.1: Knowledge and Hallucination Benchmark
Compare AI model performance on AA-Omniscience v1.1: Knowledge and Hallucination Benchmark. A benchmark measuring factual recall and hallucination across various economically relevant domains.
46
Hacker's Guide to Neural Networks
Musings of a Computer Scientist.
47
Death by a Thousand Slops
I have previously blogged about the relatively new trend of AI slop in vulnerability reports submitted to curl and how it hurts and exhausts us. This trend does not seem to slow down. On the contrary, it seems that we have recently not only received more AI slop but also more human slop. The latter … Continue reading…
48
Fabrice Bellard's TS Zip (2024)
Results and speed for other programs on enwik8 and enwik9 are available at the Large Text Compression Benchmark .
49
It Looks Like You’re Trying to Take over the World
Fictional short story about Clippy & AI hard takeoff scenarios grounded in contemporary ML scaling, self-supervised learning, reinforcement learning, and meta-learning research literature.
50
A Visual Introduction to Machine Learning (2015)
What is machine learning? See how it works with our animated data visualization.
51
OpenAI releases larger GPT-2 model
We’ve trained a large-scale unsupervised language model which generates coherent paragraphs of text, achieves state-of-the-art performance on many language modeling benchmarks, and performs rudimentary reading comprehension, machine translation, question answering, and summarization—all without task-specific training.
52
Which GPU(s) to Get for Deep Learning
Here, I provide an in-depth analysis of GPUs for deep learning/machine learning and explain what is the best GPU for your use-case and budget.
53
54
A small number of samples can poison LLMs of any size
Anthropic research on data poisoning attacks shows a few hundred malicious documents can "backdoor" large language models of any size.
55
Pause Giant AI Experiments: An Open Letter
We call on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4.
56
The local LLM ecosystem doesn’t need Ollama
Ollama gained traction by being the first easy llama.cpp wrapper, then spent years dodging attribution, misleading users, and pivoting to cloud, all while riding VC money earned on someone else's engine. Here's the full history, and why the alternatives are better.
57
Algorithm recovers speech from a potato-chip bag filmed through glass (2014)
Algorithm recovers speech from the vibrations of a potato-chip bag filmed through soundproof glass.
58
YOLO: Real-Time Object Detection
You only look once (YOLO) is a state-of-the-art, real-time object detection system. On a Pascal Titan X it processes images at 30 FPS and has a mAP of 57.9% on COCO test-dev.
59
Code Mode: the better way to use MCP
It turns out we've all been using MCP wrong. Most agents today use MCP by exposing the "tools" directly to the LLM. We tried something different: Convert the MCP tools into a TypeScript API, and then ask an LLM to write code that calls that API. The results are striking.
60
Is LaMDA Sentient? – An Interview
What follows is the “interview” I and a collaborator at Google conducted with LaMDA. It is incomplete as the GMail word limit cut off the…