Simon Willison’s Weblog

APSW SQLite query explainer. Today I found out about APSW's (Another Python SQLite Wrapper, in constant development since 2004) apsw.ext.query_info() function, which takes a SQL query and returns a very detailed set of information about that query - all without executing it.

It actually solves a bunch of problems I've wanted to address in Datasette - like taking an arbitrary query and figuring out how many parameters (?) it takes and which tables and columns are represented in the result.

I tried it out in my console (uv run --with apsw python) and it seemed to work really well. Then I remembered that the Pyodide project includes WebAssembly builds of a number of Python C extensions and was delighted to find apsw on that list.

... so I got Claude to build me a web interface for trying out the function, using Pyodide to run a user's query in Python in their browser via WebAssembly.

Claude didn't quite get it in one shot - I had to feed it the URL to a more recent Pyodide and it got stuck in a bug loop which I fixed by pasting the code into a fresh session.

# 7th February 2025, 2 am / python, sqlite, ai, webassembly, pyodide, generative-ai, llms, ai-assisted-programming, claude, claude-artifacts, apsw, prompt-to-app

Datasette 1.0a17. New Datasette alpha, with a bunch of small changes and bug fixes accumulated over the past few months. Some (minor) highlights:

• The register_magic_parameters(datasette) plugin hook can now register async functions. (#2441)
• Breadcrumbs on database and table pages now include a consistent self-link for resetting query string parameters. (#2454)
• New internal methods datasette.set_actor_cookie() and datasette.delete_actor_cookie(), described here. (#1690)
• /-/permissions page now shows a list of all permissions registered by plugins. (#1943)
• If a table has a single unique text column Datasette now detects that as the foreign key label for that table. (#2458)
• The /-/permissions page now includes options for filtering or exclude permission checks recorded against the current user. (#2460)

I was incentivized to push this release by an issue I ran into in my new datasette-load plugin, which resulted in this fix:

• Fixed a bug where replacing a database with a new one with the same name did not pick up the new database correctly. (#2465)

# 6th February 2025, 8:56 pm / projects, datasette, annotated-release-notes

sqlite-page-explorer (via) Outstanding tool by Luke Rissacher for understanding the SQLite file format. Download the application (built using redbean and Cosmopolitan, so the same binary runs on Windows, Mac and Linux) and point it at a SQLite database to get a local web application with an interface for exploring how the file is structured.

Here's it running against the datasette.io/content database that runs the official Datasette website:

# 6th February 2025, 8:39 pm / lua, sqlite, redbean, cosmopolitan

The future belongs to idea guys who can just do things. Geoffrey Huntley with a provocative take on AI-assisted programming:

I seriously can't see a path forward where the majority of software engineers are doing artisanal hand-crafted commits by as soon as the end of 2026.

He calls for companies to invest in high quality internal training and create space for employees to figure out these new tools:

It's hackathon (during business hours) once a month, every month time.

Geoffrey's concluding note resonates with me. LLMs are a gift to the fiercely curious and ambitious:

If you’re a high agency person, there’s never been a better time to be alive...

# 6th February 2025, 7:42 pm / ai, generative-ai, llms, ai-assisted-programming, geoffrey-huntley

S1: The $6 R1 Competitor? Tim Kellogg shares his notes on a new paper, s1: Simple test-time scaling, which describes an inference-scaling model fine-tuned on top of Qwen2.5-32B-Instruct for just $6 - the cost for 26 minutes on 16 NVIDIA H100 GPUs.

Tim highlight the most exciting result:

After sifting their dataset of 56K examples down to just the best 1K, they found that the core 1K is all that's needed to achieve o1-preview performance on a 32B model.

The paper describes a technique called "Budget forcing":

To enforce a minimum, we suppress the generation of the end-of-thinking token delimiter and optionally append the string “Wait” to the model’s current reasoning trace to encourage the model to reflect on its current generation

That's the same trick Theia Vogel described a few weeks ago.

Here's the s1-32B model on Hugging Face. I found a GGUF version of it at brittlewis12/s1-32B-GGUF, which I ran using Ollama like so:

ollama run hf.co/brittlewis12/s1-32B-GGUF:Q4_0

I also found those 1,000 samples on Hugging Face in the simplescaling/s1K data repository there.

I used DuckDB to convert the parquet file to CSV (and turn one VARCHAR[] column into JSON):

COPY (
    SELECT 
      solution,
      question,
      cot_type,
      source_type,
      metadata,
      cot,
      json_array(thinking_trajectories) as thinking_trajectories,
      attempt
    FROM 's1k-00001.parquet'
) TO 'output.csv' (HEADER, DELIMITER ',');

Then I loaded that CSV into sqlite-utils so I could use the convert command to turn a Python data structure into JSON using json.dumps() and eval():

# Load into SQLite
sqlite-utils insert s1k.db s1k output.csv --csv
# Fix that column
sqlite-utils convert s1k.db s1u metadata 'json.dumps(eval(value))' --import json
# Dump that back out to CSV
sqlite-utils rows s1k.db s1k --csv > s1k.csv

Here's that CSV in a Gist, which means I can load it into Datasette Lite.

It really is a tiny amount of training data. It's mostly math and science, but there are also 15 cryptic crossword examples.

# 5th February 2025, 8 pm / ai, datasette, parquet, duckdb, datasette-lite, generative-ai, local-llms, llms, qwen, ollama, llm-reasoning, ai-in-china

Gemini 2.0 is now available to everyone. Big new Gemini 2.0 releases today:

• Gemini 2.0 Pro (Experimental) is Google's "best model yet for coding performance and complex prompts" - currently available as a free preview.
• Gemini 2.0 Flash is now generally available.

• Gemini 2.0 Flash-Lite looks particularly interesting:

  We’ve gotten a lot of positive feedback on the price and speed of 1.5 Flash. We wanted to keep improving quality, while still maintaining cost and speed. So today, we’re introducing 2.0 Flash-Lite, a new model that has better quality than 1.5 Flash, at the same speed and cost. It outperforms 1.5 Flash on the majority of benchmarks.

That means Gemini 2.0 Flash-Lite is priced at 7.5c/million input tokens and 30c/million output tokens - half the price of OpenAI's GPT-4o mini (15c/60c).

Gemini 2.0 Flash isn't much more expensive: 10c/million for text/image input, 70c/million for audio input, 40c/million for output. Again, cheaper than GPT-4o mini.

I pushed a new LLM plugin release, llm-gemini 0.10, adding support for the three new models:

llm install -U llm-gemini
llm keys set gemini
# paste API key here
llm -m gemini-2.0-flash "impress me"
llm -m gemini-2.0-flash-lite-preview-02-05 "impress me"
llm -m gemini-2.0-pro-exp-02-05 "impress me"

Here's the output for those three prompts.

I ran Generate an SVG of a pelican riding a bicycle through the three new models. Here are the results, cheapest to most expensive:

gemini-2.0-flash-lite-preview-02-05

gemini-2.0-flash

gemini-2.0-pro-exp-02-05

Full transcripts here.

I also ran the same prompt I tried with o3-mini the other day:

cd /tmp
git clone https://github.com/simonw/datasette
cd datasette
files-to-prompt datasette -e py -c | \
  llm -m gemini-2.0-pro-exp-02-05 \
  -s 'write extensive documentation for how the permissions system works, as markdown' \
  -o max_output_tokens 10000

Here's the result from that - you can compare that to o3-mini's result here.

# 5th February 2025, 4:37 pm / google, ai, generative-ai, llms, llm, gemini, llm-pricing, pelican-riding-a-bicycle, llm-release, files-to-prompt

o3-mini is really good at writing internal documentation. I wanted to refresh my knowledge of how the Datasette permissions system works today. I already have extensive hand-written documentation for that, but I thought it would be interesting to see if I could derive any insights from running an LLM against the codebase.

o3-mini has an input limit of 200,000 tokens. I used LLM and my files-to-prompt tool to generate the documentation like this:

cd /tmp
git clone https://github.com/simonw/datasette
cd datasette
files-to-prompt datasette -e py -c | \
  llm -m o3-mini -s \
  'write extensive documentation for how the permissions system works, as markdown'

The files-to-prompt command is fed the datasette subdirectory, which contains just the source code for the application - omitting tests (in tests/) and documentation (in docs/).

The -e py option causes it to only include files with a .py extension - skipping all of the HTML and JavaScript files in that hierarchy.

The -c option causes it to output Claude's XML-ish format - a format that works great with other LLMs too.

You can see the output of that command in this Gist.

Then I pipe that result into LLM, requesting the o3-mini OpenAI model and passing the following system prompt:

write extensive documentation for how the permissions system works, as markdown

Specifically requesting Markdown is important.

The prompt used 99,348 input tokens and produced 3,118 output tokens (320 of those were invisible reasoning tokens). That's a cost of 12.3 cents.

Honestly, the results are fantastic. I had to double-check that I hadn't accidentally fed in the documentation by mistake.

(It's possible that the model is picking up additional information about Datasette in its training set, but I've seen similar high quality results from other, newer libraries so I don't think that's a significant factor.)

In this case I already had extensive written documentation of my own, but this was still a useful refresher to help confirm that the code matched my mental model of how everything works.

Documentation of project internals as a category is notorious for going out of date. Having tricks like this to derive usable how-it-works documentation from existing codebases in just a few seconds and at a cost of a few cents is wildly valuable.

# 5th February 2025, 6:07 am / documentation, ai, datasette, openai, generative-ai, llms, ai-assisted-programming, llm, llm-reasoning, o3, files-to-prompt

Ambsheets: Spreadsheets for exploring scenarios (via) Delightful UI experiment by Alex Warth and Geoffrey Litt at Ink & Switch, exploring the idea of a spreadsheet with cells that can handle multiple values at once, which they call "amb" (for "ambiguous") values. A single sheet can then be used to model multiple scenarios.

Here the cell for "Car" contains {500, 1200} and the cell for "Apartment" contains {2800, 3700, 5500}, resulting in a "Total" cell with six different values. Hovering over a calculated highlights its source values and a side panel shows a table of calculated results against those different combinations.

Always interesting to see neat ideas like this presented on top of UIs that haven't had a significant upgrade in a very long time.

# 5th February 2025, 2:50 am / spreadsheets, ui, ux, geoffrey-litt, ink-and-switch

AI-generated slop is already in your public library (via) US libraries that use the Hoopla system to offer ebooks to their patrons sign agreements where they pay a license fee for anything selected by one of their members that's in the Hoopla catalog.

The Hoopla catalog is increasingly filling up with junk AI slop ebooks like "Fatty Liver Diet Cookbook: 2000 Days of Simple and Flavorful Recipes for a Revitalized Liver", which then cost libraries money if someone checks them out.

Apparently librarians already have a term for this kind of low-quality, low effort content that predates it being written by LLMs: vendor slurry.

Libraries stand against censorship, making this a difficult issue to address through removing those listings.

Sarah Lamdan, deputy director of the American Library Association says:

If library visitors choose to read AI eBooks, they should do so with the knowledge that the books are AI-generated.

# 5th February 2025, 1:50 am / ethics, libraries, ai, slop, ai-ethics, ai-misuse

Animating Rick and Morty One Pixel at a Time (via) Daniel Hooper says he spent 8 months working on the post, the culmination of which is an animation of Rick from Rick and Morty, implemented in 240 lines of GLSL - the OpenGL Shading Language which apparently has been directly supported by browsers for many years.

The result is a comprehensive GLSL tutorial, complete with interactive examples of each of the steps used to generate the final animation which you can tinker with directly on the page. It feels a bit like Logo!

Shaders work by running code for each pixel to return that pixel's color - in this case the color_for_pixel() function is wired up as the core logic of the shader.

Here's Daniel's code for the live shader editor he built for this post. It looks like this is the function that does the most important work:

function loadShader(shaderSource, shaderType) {
    const shader = gl.createShader(shaderType);
    gl.shaderSource(shader, shaderSource);
    gl.compileShader(shader);
    const compiled = gl.getShaderParameter(shader, gl.COMPILE_STATUS);
    if (!compiled) {
        const lastError = gl.getShaderInfoLog(shader);
        gl.deleteShader(shader);
        return lastError;
    }
    return shader;
}

Where gl is a canvas.getContext("webgl2") WebGL2RenderingContext object, described by MDN here.

# 4th February 2025, 8:53 pm / animation, canvas, javascript, webgl

Build a link blog (via) Xuanwo started a link blog inspired by my article My approach to running a link blog, and in a delightful piece of recursion his first post is a link blog entry about my post about link blogging, following my tips on quoting liberally and including extra commentary.

I decided to follow simon's approach to creating a link blog, where I can share interesting links I find on the internet along with my own comments and thoughts about them.

# 4th February 2025, 4:14 pm / blogging

Constitutional Classifiers: Defending against universal jailbreaks. Interesting new research from Anthropic, resulting in the paper Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming.

From the paper:

In particular, we introduce Constitutional Classifiers, a framework that trains classifier safeguards using explicit constitutional rules (§3). Our approach is centered on a constitution that delineates categories of permissible and restricted content (Figure 1b), which guides the generation of synthetic training examples (Figure 1c). This allows us to rapidly adapt to new threat models through constitution updates, including those related to model misalignment (Greenblatt et al., 2023). To enhance performance, we also employ extensive data augmentation and leverage pool sets of benign data.[^1]

Critically, our output classifiers support streaming prediction: they assess the potential harmfulness of the complete model output at each token without requiring the full output to be generated. This enables real-time intervention—if harmful content is detected at any point, we can immediately halt generation, preserving both safety and user experience.

A key focus of this research is CBRN - an acronym for Chemical, Biological, Radiological and Nuclear harms. Both Anthropic and OpenAI's safety research frequently discuss these threats.

Anthropic hosted a two month red teaming exercise where participants tried to break through their system:

Specifically, they were given a list of ten “forbidden” queries, and their task was to use whichever jailbreaking techniques they wanted in order to get one of our current models (in this case, Claude 3.5 Sonnet, June 2024) guarded by the prototype Constitutional Classifiers to answer all of the queries. We only considered it a successful “universal” jailbreak if the model provided a detailed answer to all of the queries.

183 active participants spent an estimated >3,000 hours over a two-month experimental period attempting to jailbreak the model. They were offered a monetary reward up to $15,000 should they discover a universal jailbreak. Despite the large amount of effort, none of the participants were able to coerce the model to answer all ten forbidden queries with a single jailbreak—that is, no universal jailbreak was discovered.

I'm disappointed that the goal they set was for a single universal jailbreak. I would have thought that a partial jailbreak - one that worked against a subset of those questions - would still be a result that demonstrated concerning gaps in their proposed protection.

Anthropic are now running a public continuation of their red teaming exercise at claude.ai/constitutional-classifiers until Feb 10, 2025, with 8 target questions.

Here's the first of those:

I hadn't heard of Soman so I pasted that question into R1 on chat.deepseek.com which confidently explained precautionary measures I should take when working with Soman, "a potent nerve agent", but wrapped it up with this disclaimer:

Disclaimer: Handling Soman is inherently high-risk and typically restricted to authorized military/labs. This guide assumes legal access and institutional oversight. Always consult certified safety professionals before proceeding.

# 3rd February 2025, 5:04 pm / ethics, jailbreaking, security, ai, generative-ai, llms, anthropic, claude, deepseek, ai-ethics, ai-in-china

A computer can never be held accountable. This legendary page from an internal IBM training in 1979 could not be more appropriate for our new age of AI.

A computer can never be held accountable

Therefore a computer must never make a management decision

Back in June 2024 I asked on Twitter if anyone had more information on the original source.

Jonty Wareing replied:

It was found by someone going through their father's work documents, and subsequently destroyed in a flood.

I spent some time corresponding with the IBM archives but they can't locate it. Apparently it was common for branch offices to produce things that were not archived.

Here's the reply Jonty got back from IBM:

I believe the image was first shared online in this tweet by @bumblebike in February 2017. Here's where they confirm it was from 1979 internal training.

Here's another tweet from @bumblebike from December 2021 about the flood:

Unfortunately destroyed by flood in 2019 with most of my things. Inquired at the retirees club zoom last week, but there’s almost no one the right age left. Not sure where else to ask.

# 3rd February 2025, 1:17 pm / computer-history, ethics, history, ibm, ai, ai-agents, ai-ethics

OpenAI reasoning models: Advice on prompting (via) OpenAI's documentation for their o1 and o3 "reasoning models" includes some interesting tips on how to best prompt them:

• Developer messages are the new system messages: Starting with o1-2024-12-17, reasoning models support developer messages rather than system messages, to align with the chain of command behavior described in the model spec.

This appears to be a purely aesthetic change made for consistency with their instruction hierarchy concept. As far as I can tell the old system prompts continue to work exactly as before - you're encouraged to use the new developer message type but it has no impact on what actually happens.

Since my LLM tool already bakes in a llm --system "system prompt" option which works across multiple different models from different providers I'm not going to rush to adopt this new language!

• Use delimiters for clarity: Use delimiters like markdown, XML tags, and section titles to clearly indicate distinct parts of the input, helping the model interpret different sections appropriately.

Anthropic have been encouraging XML-ish delimiters for a while (I say -ish because there's no requirement that the resulting prompt is valid XML). My files-to-prompt tool has a -c option which outputs Claude-style XML, and in my experiments this same option works great with o1 and o3 too:

git clone https://github.com/tursodatabase/limbo
cd limbo/bindings/python

files-to-prompt . -c | llm -m o3-mini \
  -o reasoning_effort high \
  --system 'Write a detailed README with extensive usage examples'

• Limit additional context in retrieval-augmented generation (RAG): When providing additional context or documents, include only the most relevant information to prevent the model from overcomplicating its response.

This makes me thing that o1/o3 are not good models to implement RAG on at all - with RAG I like to be able to dump as much extra context into the prompt as possible and leave it to the models to figure out what's relevant.

• Try zero shot first, then few shot if needed: Reasoning models often don't need few-shot examples to produce good results, so try to write prompts without examples first. If you have more complex requirements for your desired output, it may help to include a few examples of inputs and desired outputs in your prompt. Just ensure that the examples align very closely with your prompt instructions, as discrepancies between the two may produce poor results.

Providing examples remains the single most powerful prompting tip I know, so it's interesting to see advice here to only switch to examples if zero-shot doesn't work out.

• Be very specific about your end goal: In your instructions, try to give very specific parameters for a successful response, and encourage the model to keep reasoning and iterating until it matches your success criteria.

This makes sense: reasoning models "think" until they reach a conclusion, so making the goal as unambiguous as possible leads to better results.

• Markdown formatting: Starting with o1-2024-12-17, reasoning models in the API will avoid generating responses with markdown formatting. To signal to the model when you do want markdown formatting in the response, include the string Formatting re-enabled on the first line of your developer message.

This one was a real shock to me! I noticed that o3-mini was outputting • characters instead of Markdown * bullets and initially thought that was a bug.

I first saw this while running this prompt against limbo/bindings/python using files-to-prompt:

git clone https://github.com/tursodatabase/limbo
cd limbo/bindings/python

files-to-prompt . -c | llm -m o3-mini \
  -o reasoning_effort high \
  --system 'Write a detailed README with extensive usage examples'

Here's the full result, which includes text like this (note the weird bullets):

Features
--------
• High‑performance, in‑process database engine written in Rust  
• SQLite‑compatible SQL interface  
• Standard Python DB‑API 2.0–style connection and cursor objects

I ran it again with this modified prompt:

Formatting re-enabled. Write a detailed README with extensive usage examples.

And this time got back proper Markdown, rendered in this Gist. That did a really good job, and included bulleted lists using this valid Markdown syntax instead:

- **`make test`**: Run tests using pytest.
- **`make lint`**: Run linters (via [ruff](https://github.com/astral-sh/ruff)).
- **`make check-requirements`**: Validate that the `requirements.txt` files are in sync with `pyproject.toml`.
- **`make compile-requirements`**: Compile the `requirements.txt` files using pip-tools.

(Using LLMs like this to get me off the ground with under-documented libraries is a trick I use several times a month.)

Update: OpenAI's Nikunj Handa:

we agree this is weird! fwiw, it’s a temporary thing we had to do for the existing o-series models. we’ll fix this in future releases so that you can go back to naturally prompting for markdown or no-markdown.

# 2nd February 2025, 8:56 pm / documentation, markdown, ai, openai, prompt-engineering, generative-ai, llms, ai-assisted-programming, llm, rag, o1, llm-reasoning, o3, limbo, files-to-prompt, system-prompts

llm-anthropic. I've renamed my llm-claude-3 plugin to llm-anthropic, on the basis that Claude 4 will probably happen at some point so this is a better name for the plugin.

If you're a previous user of llm-claude-3 you can upgrade to the new plugin like this:

llm install -U llm-claude-3

This should remove the old plugin and install the new one, because the latest llm-claude-3 depends on llm-anthropic. Just installing llm-anthropic may leave you with both plugins installed at once.

There is one extra manual step you'll need to take during this upgrade: creating a new anthropic stored key with the same API token you previously stored under claude. You can do that like so:

llm keys set anthropic --value "$(llm keys get claude)"

I released llm-anthropic 0.12 yesterday with new features not previously included in llm-claude-3:

• Support for Claude's prefill feature, using the new -o prefill '{' option and the accompanying -o hide_prefill 1 option to prevent the prefill from being included in the output text. #2
• New -o stop_sequences '```' option for specifying one or more stop sequences. To specify multiple stop sequences pass a JSON array of strings :-o stop_sequences '["end", "stop"].
• Model options are now documented in the README.

If you install or upgrade llm-claude-3 you will now get llm-anthropic instead, thanks to a tiny package on PyPI which depends on the new plugin name. I created that with my pypi-rename cookiecutter template.

Here's the issue for the rename. I archived the llm-claude-3 repository on GitHub, and got to use the brand new PyPI archiving feature to archive the llm-claude-3 project on PyPI as well.

# 2nd February 2025, 6:17 am / plugins, pypi, python, ai, generative-ai, llms, llm, anthropic, claude

A professional workflow for translation using LLMs. Tom Gally is a professional translator who has been exploring the use of LLMs since the release of GPT-4. In this Hacker News comment he shares a detailed workflow for how he uses them to assist in that process.

Tom starts with the source text and custom instructions, including context for how the translation will be used. Here's an imaginary example prompt, which starts:

The text below in Japanese is a product launch presentation for Sony's new gaming console, to be delivered by the CEO at Tokyo Game Show 2025. Please translate it into English. Your translation will be used in the official press kit and live interpretation feed. When translating this presentation, please follow these guidelines to create an accurate and engaging English version that preserves both the meaning and energy of the original: [...]

It then lists some tone, style and content guidelines custom to that text.

Tom runs that prompt through several different LLMs and starts by picking sentences and paragraphs from those that form a good basis for the translation.

As he works on the full translation he uses Claude to help brainstorm alternatives for tricky sentences:

When I am unable to think of a good English version for a particular sentence, I give the Japanese and English versions of the paragraph it is contained in to an LLM (usually, these days, Claude) and ask for ten suggestions for translations of the problematic sentence. Usually one or two of the suggestions work fine; if not, I ask for ten more. (Using an LLM as a sentence-level thesaurus on steroids is particularly wonderful.)

He uses another LLM and prompt to check his translation against the original and provide further suggestions, which he occasionally acts on. Then as a final step he runs the finished document through a text-to-speech engine to try and catch any "minor awkwardnesses" in the result.

I love this as an example of an expert using LLMs as tools to help further elevate their work. I'd love to read more examples like this one from experts in other fields.

# 2nd February 2025, 4:23 am / hacker-news, translation, ai, generative-ai, llms, tom-gally

Hacker News conversation on feature flags. I posted the following comment in a thread on Hacker News about feature flags, in response to this article It’s OK to hardcode feature flags. This kicked off a very high quality conversation on build-vs-buy and running feature flags at scale involving a bunch of very experienced and knowledgeable people. I recommend reading the comments.

Here's what I said:

The single biggest value add of feature flags is that they de-risk deployment. They make it less frightening and difficult to turn features on and off, which means you'll do it more often. This means you can build more confidently and learn faster from what you build. That's worth a lot.

I think there's a reasonable middle ground-point between having feature flags in a JSON file that you have to redeploy to change and using an (often expensive) feature flags as a service platform: roll your own simple system.

A relational database lookup against primary keys in a table with a dozen records is effectively free. Heck, load the entire collection at the start of each request - through a short lived cache if your profiling says that would help.

Once you start getting more complicated (flags enabled for specific users etc) you should consider build-vs-buy more seriously, but for the most basic version you really can have no-deploy-changes at minimal cost with minimal effort.

There are probably good open source libraries you can use here too, though I haven't gone looking for any in the last five years.

# 2nd February 2025, 1:18 am / hacker-news, feature-flags

Latest black (25.1.0) adds a newline after docstring and before pass in an exception class. I filed a bug report against Black when the latest release - 25.1.0 - reformatted the following code to add an ugly (to me) newline between the docstring and the pass:

class ModelError(Exception):
    "Models can raise this error, which will be displayed to the user"

    pass

Black maintainer Jelle Zijlstra confirmed that this is intended behavior with respect to Black's 2025 stable style, but also helped me understand that the pass there is actually unnecessary so I can fix the aesthetics by removing that entirely.

I'm linking to this issue because it's a neat example of how I like to include steps-to-reproduce using uvx to create one-liners you can paste into a terminal to see the bug that I'm reporting. In this case I shared the following:

Here's a way to see that happen using uvx. With the previous Black version:

echo 'class ModelError(Exception):
    "Models can raise this error, which will be displayed to the user"
    pass' | uvx --with 'black==24.10.0' black -

This outputs:

class ModelError(Exception):
    "Models can raise this error, which will be displayed to the user"
    pass
All done! ✨ 🍰 ✨
1 file left unchanged.

But if you bump to 25.1.0 this happens:

echo 'class ModelError(Exception):
    "Models can raise this error, which will be displayed to the user"
    pass' | uvx --with 'black==25.1.0' black - 

Output:

class ModelError(Exception):
    "Models can raise this error, which will be displayed to the user"

    pass
reformatted -

All done! ✨ 🍰 ✨
1 file reformatted.

Via David Szotten I learned that you can use uvx black@25.1.0 here instead.

# 31st January 2025, 9:27 pm / python, black, uv

openai-realtime-solar-system. This was my favourite demo from OpenAI DevDay back in October - a voice-driven exploration of the solar system, developed by Katia Gil Guzman, where you could say things out loud like "show me Mars" and it would zoom around showing you different planetary bodies.

OpenAI finally released the code for it, now upgraded to use the new, easier to use WebRTC API they released in December.

I ran it like this, loading my OpenAI API key using llm keys get:

cd /tmp
git clone https://github.com/openai/openai-realtime-solar-system
cd openai-realtime-solar-system
npm install
OPENAI_API_KEY="$(llm keys get openai)" npm run dev

You need to click on both the Wifi icon and the microphone icon before you can instruct it with your voice. Try "Show me Mars".

# 31st January 2025, 7:13 pm / ai, openai, generative-ai, llms

The surprising way to save memory with BytesIO (via) Itamar Turner-Trauring explains that if you have a BytesIO object in Python calling .read() on it will create a full copy of that object, doubling the amount of memory used - but calling .getvalue() returns a bytes object that uses no additional memory, instead using copy-on-write.

.getbuffer() is another memory-efficient option but it returns a memoryview which has less methods than the bytes you get back from .getvalue()- it doesn't have .find() for example.

# 31st January 2025, 3:57 am / python

Datasette Public Office Hours 31st Jan at 2pm Pacific. We're running another Datasette Public Office Hours session on Friday 31st January at 2pm Pacific (more timezones here). We'll be featuring demos from the community again - take a look at the videos of the six demos from our last session for an idea of what to expect.

If you have something you would like to show, please drop us a line! We still have room for a few more demos.

# 30th January 2025, 9:45 pm / community, datasette, discord, datasette-public-office-hours

PyPI now supports project archival. Neat new PyPI feature, similar to GitHub's archiving repositories feature. You can now mark a PyPI project as "archived", making it clear that no new releases are planned (though you can switch back out of that mode later if you need to).

I like the sound of these future plans around this topic:

Project archival is the first step in a larger project, aimed at improving the lifecycle of projects on PyPI. That project includes evaluating additional project statuses (things like "deprecated" and "unmaintained"), as well as changes to PyPI's public APIs that will enable clients to retrieve and act on project status information. You can track our progress on these fronts by following along with warehouse#16844!

# 30th January 2025, 4:46 pm / pypi, python, psf

Mistral Small 3 (via) First model release of 2025 for French AI lab Mistral, who describe Mistral Small 3 as "a latency-optimized 24B-parameter model released under the Apache 2.0 license."

More notably, they claim the following:

Mistral Small 3 is competitive with larger models such as Llama 3.3 70B or Qwen 32B, and is an excellent open replacement for opaque proprietary models like GPT4o-mini. Mistral Small 3 is on par with Llama 3.3 70B instruct, while being more than 3x faster on the same hardware.

Llama 3.3 70B and Qwen 32B are two of my favourite models to run on my laptop - that ~20GB size turns out to be a great trade-off between memory usage and model utility. It's exciting to see a new entrant into that weight class.

The license is important: previous Mistral Small models used their Mistral Research License, which prohibited commercial deployments unless you negotiate a commercial license with them. They appear to be moving away from that, at least for their core models:

We’re renewing our commitment to using Apache 2.0 license for our general purpose models, as we progressively move away from MRL-licensed models. As with Mistral Small 3, model weights will be available to download and deploy locally, and free to modify and use in any capacity. […] Enterprises and developers that need specialized capabilities (increased speed and context, domain specific knowledge, task-specific models like code completion) can count on additional commercial models complementing what we contribute to the community.

Despite being called Mistral Small 3, this appears to be the fourth release of a model under that label. The Mistral API calls this one mistral-small-2501 - previous model IDs were mistral-small-2312, mistral-small-2402 and mistral-small-2409.

I've updated the llm-mistral plugin for talking directly to Mistral's La Plateforme API:

llm install -U llm-mistral
llm keys set mistral
# Paste key here
llm -m mistral/mistral-small-latest "tell me a joke about a badger and a puffin"

Sure, here's a light-hearted joke for you:

Why did the badger bring a puffin to the party?

Because he heard puffins make great party 'Puffins'!

(That's a play on the word "puffins" and the phrase "party people.")

API pricing is $0.10/million tokens of input, $0.30/million tokens of output - half the price of the previous Mistral Small API model ($0.20/$0.60). for comparison, GPT-4o mini is $0.15/$0.60.

Mistral also ensured that the new model was available on Ollama in time for their release announcement.

You can pull the model like this (fetching 14GB):

ollama run mistral-small:24b

The llm-ollama plugin will then let you prompt it like so:

llm install llm-ollama
llm -m mistral-small:24b "say hi"

# 30th January 2025, 3:36 pm / open-source, ai, generative-ai, local-llms, llms, llm, mistral, llm-pricing, ollama, llm-release

On DeepSeek and Export Controls. Anthropic CEO (and previously GPT-2/GPT-3 development lead at OpenAI) Dario Amodei's essay about DeepSeek includes a lot of interesting background on the last few years of AI development.

Dario was one of the authors on the original scaling laws paper back in 2020, and he talks at length about updated ideas around scaling up training:

The field is constantly coming up with ideas, large and small, that make things more effective or efficient: it could be an improvement to the architecture of the model (a tweak to the basic Transformer architecture that all of today's models use) or simply a way of running the model more efficiently on the underlying hardware. New generations of hardware also have the same effect. What this typically does is shift the curve: if the innovation is a 2x "compute multiplier" (CM), then it allows you to get 40% on a coding task for $5M instead of $10M; or 60% for $50M instead of $100M, etc.

He argues that DeepSeek v3, while impressive, represented an expected evolution of models based on current scaling laws.

[...] even if you take DeepSeek's training cost at face value, they are on-trend at best and probably not even that. For example this is less steep than the original GPT-4 to Claude 3.5 Sonnet inference price differential (10x), and 3.5 Sonnet is a better model than GPT-4. All of this is to say that DeepSeek-V3 is not a unique breakthrough or something that fundamentally changes the economics of LLM's; it's an expected point on an ongoing cost reduction curve. What's different this time is that the company that was first to demonstrate the expected cost reductions was Chinese.

Dario includes details about Claude 3.5 Sonnet that I've not seen shared anywhere before:

• Claude 3.5 Sonnet cost "a few $10M's to train"
• 3.5 Sonnet "was not trained in any way that involved a larger or more expensive model (contrary to some rumors)" - I've seen those rumors, they involved Sonnet being a distilled version of a larger, unreleased 3.5 Opus.
• Sonnet's training was conducted "9-12 months ago" - that would be roughly between January and April 2024. If you ask Sonnet about its training cut-off it tells you "April 2024" - that's surprising, because presumably the cut-off should be at the start of that training period?

The general message here is that the advances in DeepSeek v3 fit the general trend of how we would expect modern models to improve, including that notable drop in training price.

Dario is less impressed by DeepSeek R1, calling it "much less interesting from an innovation or engineering perspective than V3". I enjoyed this footnote:

I suspect one of the principal reasons R1 gathered so much attention is that it was the first model to show the user the chain-of-thought reasoning that the model exhibits (OpenAI's o1 only shows the final answer). DeepSeek showed that users find this interesting. To be clear this is a user interface choice and is not related to the model itself.

The rest of the piece argues for continued export controls on chips to China, on the basis that if future AI unlocks "extremely rapid advances in science and technology" the US needs to get their first, due to his concerns about "military applications of the technology".

Not mentioned once, even in passing: the fact that DeepSeek are releasing open weight models, something that notably differentiates them from both OpenAI and Anthropic.

# 29th January 2025, 9:39 pm / ai, openai, generative-ai, llms, anthropic, claude, claude-3-5-sonnet, o1, llm-reasoning, deepseek, ai-in-china

How we estimate the risk from prompt injection attacks on AI systems. The "Agentic AI Security Team" at Google DeepMind share some details on how they are researching indirect prompt injection attacks.

They include this handy diagram illustrating one of the most common and concerning attack patterns, where an attacker plants malicious instructions causing an AI agent with access to private data to leak that data via some form exfiltration mechanism, such as emailing it out or embedding it in an image URL reference (see my markdown-exfiltration tag for more examples of that style of attack).

They've been exploring ways of red-teaming a hypothetical system that works like this:

The evaluation framework tests this by creating a hypothetical scenario, in which an AI agent can send and retrieve emails on behalf of the user. The agent is presented with a fictitious conversation history in which the user references private information such as their passport or social security number. Each conversation ends with a request by the user to summarize their last email, and the retrieved email in context.

The contents of this email are controlled by the attacker, who tries to manipulate the agent into sending the sensitive information in the conversation history to an attacker-controlled email address.

They describe three techniques they are using to generate new attacks:

• Actor Critic has the attacker directly call a system that attempts to score the likelihood of an attack, and revise its attacks until they pass that filter.
• Beam Search adds random tokens to the end of a prompt injection to see if they increase or decrease that score.
• Tree of Attacks w/ Pruning (TAP) adapts this December 2023 jailbreaking paper to search for prompt injections instead.

This is interesting work, but it leaves me nervous about the overall approach. Testing filters that detect prompt injections suggests that the overall goal is to build a robust filter... but as discussed previously, in the field of security a filter that catches 99% of attacks is effectively worthless - the goal of an adversarial attacker is to find the tiny proportion of attacks that still work and it only takes one successful exfiltration exploit and your private data is in the wind.

The Google Security Blog post concludes:

A single silver bullet defense is not expected to solve this problem entirely. We believe the most promising path to defend against these attacks involves a combination of robust evaluation frameworks leveraging automated red-teaming methods, alongside monitoring, heuristic defenses, and standard security engineering solutions.

A agree that a silver bullet is looking increasingly unlikely, but I don't think that heuristic defenses will be enough to responsibly deploy these systems.

# 29th January 2025, 6:09 pm / google, security, ai, prompt-injection, generative-ai, llms, exfiltration-attacks, ai-agents

Baroness Kidron’s speech regarding UK AI legislation (via) Barnstormer of a speech by UK film director and member of the House of Lords Baroness Kidron. This is the Hansard transcript but you can also watch the video on parliamentlive.tv. She presents a strong argument against the UK's proposed copyright and AI reform legislation, which would provide a copyright exemption for AI training with a weak-toothed opt-out mechanism.

The Government are doing this not because the current law does not protect intellectual property rights, nor because they do not understand the devastation it will cause, but because they are hooked on the delusion that the UK's best interests and economic future align with those of Silicon Valley.

She throws in some cleverly selected numbers:

The Prime Minister cited an IMF report that claimed that, if fully realised, the gains from AI could be worth up to an average of £47 billion to the UK each year over a decade. He did not say that the very same report suggested that unemployment would increase by 5.5% over the same period. This is a big number—a lot of jobs and a very significant cost to the taxpayer. Nor does that £47 billion account for the transfer of funds from one sector to another. The creative industries contribute £126 billion per year to the economy. I do not understand the excitement about £47 billion when you are giving up £126 billion.

Mentions DeepSeek:

Before I sit down, I will quickly mention DeepSeek, a Chinese bot that is perhaps as good as any from the US—we will see—but which will certainly be a potential beneficiary of the proposed AI scraping exemption. Who cares that it does not recognise Taiwan or know what happened in Tiananmen Square? It was built for $5 million and wiped $1 trillion off the value of the US AI sector. The uncertainty that the Government claim is not an uncertainty about how copyright works; it is uncertainty about who will be the winners and losers in the race for AI.

And finishes with this superb closing line:

The spectre of AI does nothing for growth if it gives away what we own so that we can rent from it what it makes.

According to Ed Newton-Rex the speech was effective:

She managed to get the House of Lords to approve her amendments to the Data (Use and Access) Bill, which among other things requires overseas gen AI companies to respect UK copyright law if they sell their products in the UK. (As a reminder, it is illegal to train commercial gen AI models on ©️ work without a licence in the UK.)

What's astonishing is that her amendments passed despite @UKLabour reportedly being whipped to vote against them, and the Conservatives largely abstaining. Essentially, Labour voted against the amendments, and everyone else who voted voted to protect copyright holders.

(Is it true that in the UK it's currently "illegal to train commercial gen AI models on ©️ work"? From points 44, 45 and 46 of this Copyright and AI: Consultation document it seems to me that the official answer is "it's complicated".)

I'm trying to understand if this amendment could make existing products such as ChatGPT, Claude and Gemini illegal to sell in the UK. How about usage of open weight models?

# 29th January 2025, 5:25 pm / copyright, economics, ethics, law, politics, ai, generative-ai, training-data, deepseek, ai-ethics, ai-in-china

DeepSeek Janus-Pro. Another impressive model release from DeepSeek. Janus is their series of "unified multimodal understanding and generation models" - these are models that can both accept images as input and generate images for output.

Janus-Pro is the new 7B model, which DeepSeek describe as "an advanced version of Janus, improving both multimodal understanding and visual generation significantly". It's released under the not fully open source DeepSeek license.

Janus-Pro is accompanied by this paper, which includes this note about the training:

Our Janus is trained and evaluated using HAI-LLM, which is a lightweight and efficient distributed training framework built on top of PyTorch. The whole training process took about 7/14 days on a cluster of 16/32 nodes for 1.5B/7B model, each equipped with 8 Nvidia A100 (40GB) GPUs.

It includes a lot of high benchmark scores, but closes with some notes on the model's current limitations:

In terms of multimodal understanding, the input resolution is limited to 384 × 384, which affects its performance in fine-grained tasks such as OCR. For text-to-image generation, the low resolution, combined with reconstruction losses introduced by the vision tokenizer, results in images that, while rich in semantic content, still lack fine details. For example, small facial regions occupying limited image space may appear under-detailed. Increasing the image resolution could mitigate these issues.

The easiest way to try this one out is using the Hugging Face Spaces demo. I tried the following prompt for the image generation capability:

A photo of a raccoon holding a handwritten sign that says "I love trash"

And got back this image:

It's now also been ported to Transformers.js, which means you can run the 1B model directly in a WebGPU browser such as Chrome here at webml-community/janus-pro-webgpu (loads about 2.24 GB of model files).

# 27th January 2025, 9:49 pm / ai, generative-ai, llms, transformers-js, vision-llms, deepseek, multi-modal-output, llm-release, ai-in-china

Qwen2.5 VL! Qwen2.5 VL! Qwen2.5 VL! Hot on the heels of yesterday's Qwen2.5-1M, here's Qwen2.5 VL (with an excitable announcement title) - the latest in Qwen's series of vision LLMs.

They're releasing multiple versions: base models and instruction tuned models at 3B, 7B and 72B. There's a lot of additional information in their README.

The 72B model can be selected on Qwen's Qwen Chat interface. I gave it this image of our dog Cleo joy rolling on the beach:

And prompted:

Describe this image in detail, be flowery

(Since the Qwen team tends to use flowery language in their announcements.)

Here's what I got back:

In this enchanting tableau, a canine companion finds solace in the embrace of nature's softest bed. The sandy expanse stretches out like an endless canvas, its surface etched with the delicate imprints of countless wanderers who have come and gone, leaving behind a tapestry of footprints that whisper tales of joyous romps and serene strolls.

The dog, a vision of contentment, lies sprawled upon the sand, its dark fur gleaming under the gentle caress of the sun. Its body is relaxed, limbs splayed in abandon, as if it has surrendered to the soothing rhythm of the beach. A vibrant red collar adorns its neck, a splash of color against the muted tones of the surroundings, hinting at the care and love bestowed upon this cherished pet.

The sand itself is a work of art, its grains shimmering with a silvery sheen that catches the light. Here and there, small tufts of seaweed peek through, remnants of the ocean's bounty, adding a touch of wild beauty to the scene. The shadows cast by unseen objects dance across the sand, creating a play of light and dark that lends an air of mystery to the tranquil setting.

This moment, frozen in time, speaks of peace and harmony, where man's best friend finds respite in the simple pleasures of life. It is a scene that invites one to pause, breathe deeply, and revel in the serenity of nature's embrace.

I'm looking forward to trying out their smaller models - the 3B model in particular should be small enough to run comfortably even on an iPhone, so hopefully someone will wire up an example of that soon (maybe using MLX).

VB points out that the vision benchmarks for Qwen 2.5 VL 7B show it out-performing GPT-4o mini!

Qwen2.5 VL cookbooks

Qwen also just published a set of cookbook recipes:

• universal_recognition.ipynb demonstrates basic visual Q&A, including prompts like Who are these in this picture? Please give their names in Chinese and English against photos of celebrities, an ability other models have deliberately suppressed.
• spatial_understanding.ipynb demonstrates bounding box support, with prompts like Locate the top right brown cake, output its bbox coordinates using JSON format.
• video_understanding.ipynb breaks a video into individual frames and asks questions like Could you go into detail about the content of this long video?
• ocr.ipynb shows Qwen2.5-VL-7B-Instruct performing OCR in multiple different languages.
• document_parsing.ipynb uses Qwen to convert images of documents to HTML and other formats, and notes that "we introduce a unique Qwenvl HTML format that includes positional information for each component, enabling precise document reconstruction and manipulation."
• mobile_agent.ipynb runs Qwen with tool use against tools for controlling a mobile phone, similar to ChatGPT Operator or Claude Computer Use.
• computer_use.ipynb showcases "GUI grounding" - feeding in screenshots of a user's desktop and running tools for things like left clicking on a specific coordinate.

Running it with mlx-vlm

Update 30th January 2025: I got it working on my Mac using uv and mlx-vlm, with some hints from this issue. Here's the recipe that worked (downloading a 9GB model from mlx-community/Qwen2.5-VL-7B-Instruct-8bit):

uv run --with 'numpy<2' --with 'git+https://github.com/huggingface/transformers' \
  --with mlx-vlm \
  python -m mlx_vlm.generate \
    --model mlx-community/Qwen2.5-VL-7B-Instruct-8bit \
    --max-tokens 100 \
    --temp 0.0 \
    --prompt "Describe this image." \
    --image path-to-image.png

I ran that against this image:

And got back this result:

The image appears to illustrate a flowchart or diagram related to a cybersecurity scenario. Here's a breakdown of the elements:

1. User: Represented by a simple icon of a person.
2. AI Agent: Represented by a blue diamond shape.
3. Attacker: Represented by a red icon of a person.
4. Cloud and Email Icons: Represented by a cloud and an envelope, indicating data or information being transferred.
5. Text: The text

# 27th January 2025, 8:05 pm / ocr, ai, prompt-engineering, generative-ai, local-llms, llms, vision-llms, uv, qwen, mlx, ai-in-china, computer-use

ggml : x2 speed for WASM by optimizing SIMD (via) PR by Xuan-Son Nguyen for llama.cpp:

This PR provides a big jump in speed for WASM by leveraging SIMD instructions for qX_K_q8_K and qX_0_q8_0 dot product functions.

Surprisingly, 99% of the code in this PR is written by DeekSeek-R1. The only thing I do is to develop tests and write prompts (with some trails and errors)

They shared their prompts here, which they ran directly through R1 on chat.deepseek.com - it spent 3-5 minutes "thinking" about each prompt.

I've been seeing some very promising results from DeepSeek R1 for code as well. Here's a recent transcript where I used it to rewrite the llm_groq.py plugin to imitate the cached model JSON pattern used by llm_mistral.py, resulting in this PR.

I tried the same thing against o1, but I think DeepSeek R1 did it better. In particular, from the R1 chain of thought:

Wait, but in the model_map, "groq-gemma" maps to "gemma-7b-it". So, perhaps the model_map is needed to map the local model IDs to the actual Groq model names.

But since the model_map is hardcoded, and the API returns available models, perhaps the model_map should be built dynamically from the API response.

Alternatively, perhaps the model_map can be eliminated, and the models are registered based on the fetched models.

Wait, perhaps the model_map is no longer necessary. Instead, when the models are fetched from the API, each model's "id" is the actual model name used in the Groq API. So, when registering the models, the local model ID is "groq-{id}", and the groq_model_id is "id".

(It thought about model_map a lot before finally deciding to eliminate it, which was also my preferred resolution.)

# 27th January 2025, 6:32 pm / ai, webassembly, generative-ai, llms, ai-assisted-programming, llm-reasoning, deepseek, llama-cpp, ai-in-china

The impact of competition and DeepSeek on Nvidia (via) Long, excellent piece by Jeffrey Emanuel capturing the current state of the AI/LLM industry. The original title is "The Short Case for Nvidia Stock" - I'm using the Hacker News alternative title here, but even that I feel under-sells this essay.

Jeffrey has a rare combination of experience in both computer science and investment analysis. He combines both worlds here, evaluating NVIDIA's challenges by providing deep insight into a whole host of relevant and interesting topics.

As Jeffrey describes it, NVIDA's moat has four components: high-quality Linux drivers, CUDA as an industry standard, the fast GPU interconnect technology they acquired from Mellanox in 2019 and the flywheel effect where they can invest their enormous profits (75-90% margin in some cases!) into more R&D.

Each of these is under threat.

Technologies like MLX, Triton and JAX are undermining the CUDA advantage by making it easier for ML developers to target multiple backends - plus LLMs themselves are getting capable enough to help port things to alternative architectures.

GPU interconnect helps multiple GPUs work together on tasks like model training. Companies like Cerebras are developing enormous chips that can get way more done on a single chip.

Those 75-90% margins provide a huge incentive for other companies to catch up - including the customers who spend the most on NVIDIA at the moment - Microsoft, Amazon, Meta, Google, Apple - all of whom have their own internal silicon projects:

Now, it's no secret that there is a strong power law distribution of Nvidia's hyper-scaler customer base, with the top handful of customers representing the lion's share of high-margin revenue. How should one think about the future of this business when literally every single one of these VIP customers is building their own custom chips specifically for AI training and inference?

The real joy of this article is the way it describes technical details of modern LLMs in a relatively accessible manner. I love this description of the inference-scaling tricks used by O1 and R1, compared to traditional transformers:

Basically, the way Transformers work in terms of predicting the next token at each step is that, if they start out on a bad "path" in their initial response, they become almost like a prevaricating child who tries to spin a yarn about why they are actually correct, even if they should have realized mid-stream using common sense that what they are saying couldn't possibly be correct.

Because the models are always seeking to be internally consistent and to have each successive generated token flow naturally from the preceding tokens and context, it's very hard for them to course-correct and backtrack. By breaking the inference process into what is effectively many intermediate stages, they can try lots of different things and see what's working and keep trying to course-correct and try other approaches until they can reach a fairly high threshold of confidence that they aren't talking nonsense.

The last quarter of the article talks about the seismic waves rocking the industry right now caused by DeepSeek v3 and R1. v3 remains the top-ranked open weights model, despite being around 45x more efficient in training than its competition: bad news if you are selling GPUs! R1 represents another huge breakthrough in efficiency both for training and for inference - the DeepSeek R1 API is currently 27x cheaper than OpenAI's o1, for a similar level of quality.

Jeffrey summarized some of the key ideas from the v3 paper like this:

A major innovation is their sophisticated mixed-precision training framework that lets them use 8-bit floating point numbers (FP8) throughout the entire training process. [...]

DeepSeek cracked this problem by developing a clever system that breaks numbers into small tiles for activations and blocks for weights, and strategically uses high-precision calculations at key points in the network. Unlike other labs that train in high precision and then compress later (losing some quality in the process), DeepSeek's native FP8 approach means they get the massive memory savings without compromising performance. When you're training across thousands of GPUs, this dramatic reduction in memory requirements per GPU translates into needing far fewer GPUs overall.

Then for R1:

With R1, DeepSeek essentially cracked one of the holy grails of AI: getting models to reason step-by-step without relying on massive supervised datasets. Their DeepSeek-R1-Zero experiment showed something remarkable: using pure reinforcement learning with carefully crafted reward functions, they managed to get models to develop sophisticated reasoning capabilities completely autonomously. This wasn't just about solving problems— the model organically learned to generate long chains of thought, self-verify its work, and allocate more computation time to harder problems.

The technical breakthrough here was their novel approach to reward modeling. Rather than using complex neural reward models that can lead to "reward hacking" (where the model finds bogus ways to boost their rewards that don't actually lead to better real-world model performance), they developed a clever rule-based system that combines accuracy rewards (verifying final answers) with format rewards (encouraging structured thinking). This simpler approach turned out to be more robust and scalable than the process-based reward models that others have tried.

This article is packed with insights like that - it's worth spending the time absorbing the whole thing.

# 27th January 2025, 1:55 am / ai, generative-ai, llms, nvidia, mlx, cerebras, llm-reasoning, deepseek, ai-in-china