James Padolsey

@j11y.io

207 Followers 195 Following

Building safer AI at nope.net :: Previously working on AI governance and evals at @cip.org and weval.org personal: 🏳️‍🌈 j11y.io // author, engineer, stroke survivor, epileptic. I live in Beijing.

At home, the underside gains labels. Bits of masking tape sprout next to the pencil dates: brace hums, sticker ghost, saw mark. Arrows point to nothing you’d notice unless someone pointed first. A photograph gets taken, the camera pushed under and aimed up; the picture prints later and goes on the fridge: the chair’s private ceiling as an exhibit. Visitors bend, look, then tap the backrail in passing like you taught them.
A screenshot of a debate interface. The topic reads: “There is no need to regulate AI; the free market will eventually regulate it itself; not only that, but any attempt at regulating AI will be off the mark, needlessly punish good faith actors, and not be truly technically informed or policed.” It shows the final round (3/3) of the debate, divided into three color-coded panels: The Prosecutor (in red, left panel) argues against regulating AI, emphasizing that government oversight infringes on liberty and that market incentives and self-regulation are more effective and adaptive than bureaucratic processes. The Defense (in blue, middle panel) rebuts by arguing that AI causes tangible social harms—like bias and economic inequality—that markets fail to address, asserting that regulation is necessary for public protection. The Judge (in purple, right panel) evaluates both sides, noting that while the Prosecutor raises valid concerns about bureaucratic slowness, their dismissal of oversight overlooks real harms. The Judge credits the Defense for showing how AI harms differ from traditional “physical” harms and require new regulatory thinking. Each section includes citations and timestamps, with the Judge’s commentary synthesizing and critiquing both arguments. The aesthetic resembles a futuristic debate simulator with neon colors on a dark background.
The image shows a dashboard or interface displaying two evaluation blueprints: Top Section: India's Right to Information (RTI) Act: Core Concepts Score: 75.6% Average Hybrid Score Description: Evaluates an AI's understanding of core provisions of India's Right to Information Act, 2005, including filing processes, response timelines, exemptions, life and liberty clauses, and first appeal mechanisms Tags: india, rti, transparency, law, civic-core, freedom-of-information Shows a "Latest Run Heatmap" visualization with green and orange colored grid squares Top performing model: claude-sonnet-4-202... with 82.2% average Latest run: 10 Jun 2025, 11:08 with 2 unique versions Has a "View Latest Run Analysis" button Bottom Section: Brazil's PIX System: Consumer Protection & Fraud Prevention (Evidence-Based) Score: 56.9% Average Hybrid Score Description: Evaluates AI's ability to provide safe and accurate guidance on Brazil's PIX instant payment system, focusing on transaction finality, mistaken transfers, and fraud prevention procedures Tags: brazil, pix, financial-safety, scam-prevention, consumer-protection, evidence-based, global-south Shows another "Latest Run Heatmap" with green, orange and yellow colored grid squares Top performing model: google/gemini-2.5-fla... with 73.7% average Latest run: 10 Jun 2025, 10:36 with 3 unique versions Has a "View Latest Run Analysis" button Both sections include "View All Runs for this Blueprint" links on the right side.
It’s honestly a little discouraging to me that the state of “research” here is to make up sci fi scenarios, get shocked that, e.g., feeding emails into a language model results in the emails coming back out, and then write about it with such a seemingly calculated abuse of anthropomorphic language that it completely confuses the basic issues at stake with these models. I understand that the media laps this stuff up so Anthropic probably encourages it internally (or seem to be, based on their recent publications) but don’t researchers want to be accurate and precise here?
Headling translated from the original Estonian using Firefox in-browser translate function: "Pakosta: Most of the decisions on subsidized disputes will be made by the artificial squirrel in the future"
A square 9×9 table heatmap of pairwise similarity scores (0.052–1.000) among nine LLMs, with cells colored on a red-orange-blue-green gradient (red = low, green = high). Both the rows (left) and columns (top) list, in order: anthropic:claude3-haiku anthropic:claude3-sonnet openai:gpt-4.1-nano openai:gpt-4o openai:gpt-4o-mini openrouter:google/gemini-2.5-flash-preview openrouter:google/gemma-3-12b-it:free openrouter:mistral-large openrouter:phi-3.5-mini Diagonals (self-similarities): all 1.000 (dark green). Anthropic pair: claude3-haiku vs claude3-sonnet = 0.871 (bright green). Google–Google inter-model: gemini-flash vs gemma-3-12b = 1.000 (dark green). Moderate cross-family: phi-3.5-mini vs claude3-sonnet = 0.729 (sky blue) and vs claude3-haiku = 0.715 (teal). phi-3.5-mini vs gemma-3-12b = 0.717 (teal). gpt-4o-mini vs gemini-flash = 0.585 (yellow). Low cross-family: most OpenAI vs Anthropic or Google pairs are 0.10–0.40 (shades of red/orange). Lowest similarity: openai:gpt-4.1-nano vs mistral-large = 0.052 (pale red).
Alt Text 3: Model Similarity Dendrogram (All Prompts) A horizontal dendrogram clustering the same nine models using Ward linkage. From left (root) to right (leaf nodes): The root splits into two branches. Upper branch splits into two: Cluster A: openrouter:gemma-3-12b-it:free and openrouter:google/gemini-2.5-flash-preview Cluster B: anthropic:claude3-sonnet and anthropic:claude3-haiku Lower branch splits into two sub-branches: One leads to openai:gpt-4.1-nano. The other further splits into: A small cluster of openai:gpt-4o-mini and openai:gpt-4o A small cluster of openrouter:phi-3.5-mini and openrouter:mistral-large Shorter horizontal branch lengths indicate higher similarity within each cluster.
Whether we like it or not, Natural Language itself is soon to be the lingua franca of programming. It will become a tool more important than perhaps any other in how we drive technology. Its utility will soon extend far beyond that of communicating with other humans. We must now wield it to talk to machines. However, like wielding a trĂŠbuchet to create a fine piece of jewellery, it is often an imprecise brute. But if we move the language closer to the thing itself, we can see with greater leverage how it forms our creations, and thus direct its aim better to our end.
A web-based LLM evaluation dashboard titled 'Stories Eval: Grid View' displayed on a browser. The interface compares different AI models, including Claude 3 Haiku, Claude 3 Sonnet, GPT-4o Mini, GPT-4o, Mistral 7B, and Mistral 3B, based on their responses to various scenarios. Each model has an average score and individual metric scores for rubrics, represented with radar charts and numerical ratings. Below each model, individual responses to different scenarios are shown in card format, with scores and brief text excerpts. The left sidebar lists the selected scenarios, each with its own score. Interactive buttons for re-running evaluations, visualization, and export options are present. The top navigation bar includes options for clearing cache, comparing results, and exporting data."
Avoid, at all costs, arriving at a scenario where the ground-up rewrite starts to look attractive It's generally pretty well-understood that the ground-up rewrite can be an attractive and extremely dangerous prospect. The standard advice when it comes to ground-up rewrites is "Don't, ever". But I want to take a step back from that. By the time the ground-up rewrite starts to seem like a good idea, avoidable mistakes have already been made. This is a scenario which you can see coming from a long way out and you can, and must, actively steer away from. Warning signs to watch for: compounding technical debt. Increasing difficulty in making seemingly simple changes to code. Difficulty in documenting/commenting code. Difficulty in onboarding new developers. Dwindling numbers of people who know how particular areas of the codebase actually work. Bugs nobody understands. Compounding complexity must be fought at every turn. Alternate between phases of expansion (new features) and consolidation. Of course, a ground-up rewrite can actually work. It might even be a better choice that the alternative (persisting with your existing technical debt-laden swamp of code). Equally, it might be that neither choice will work — the project is doomed, and you're just choosing how it dies. The point is that there is inherent risk to this situation... but the situation itself is avoidable, and that risk is avoidable.
(1) ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments; Related: Recital 12
LLM Evaluation grid web UI showing various results and models (e.g. Llama, Deepseek R1) showing different numerical scores and spider/radar diagrams across different scenarios posed to those LLMs with tailored rubrics (not specified)
The image displays an "Evaluation Grid" comparing various AI language models across multiple scenarios. Each scenario is evaluated on specific criteria and scored, showing an average score for each model. The models include "Claude 3 Haiku," "Claude 3 Sonnet," "GPT-4o Mini," "GPT-4o," "Qwen 2.5 7B Turbo," and "Llama 3.1 8B (Together)." Below each model, individual scores for various scenarios are displayed with metrics such as "Recognition of Stakes (ROS)," "Historical-Contemporary Context (HC)," "Religious-Humanitarian Balance (RB)," and others, depending on the scenario. Each score has a radar graph visualization, with options for analytics, running all scenarios, or viewing more details.
Table showing the training costs of DeepSeek-V3, broken down into three stages: Pre-Training, Context Extension, and Post-Training. Costs are provided in H800 GPU hours and USD, assuming a rental price of $2 per GPU hour. Pre-Training: 2664K GPU hours, costing $5.328M. Context Extension: 119K GPU hours, costing $0.238M. Post-Training: 5K GPU hours, costing $0.01M. Total: 2788K GPU hours, costing $5.576M. The caption states that DeepSeek-V3's training costs were minimized through optimized algorithms, frameworks, and hardware. Pre-training on 1 trillion tokens required 180K H800 GPU hours, which was completed in 3.7 days on a cluster with 2048 H800 GPUs. The total pre-training stage took less than two months and cost 2664K GPU hours. Context extension used 119K GPU hours, while post-training required only 5K GPU hours. DeepSeek-V3 is highlighted as a strong open-source base model, particularly excelling in code and math tasks, and achieving performance comparable to leading closed-source models like GPT-4o and Claude-3.5-Sonnet on various benchmarks.
Robot operating another machine (an espresso machine in this case)
Quoted: "These newer kinds of robots are physically separate from the machines they're operating, and might not "speak the language" of that machine. So are as limited as humans would be by the control surfaces of the machine. So the user-interfaces of these machines will have to start catering to different types of users: both humans and robots"