Simon Willison’s Weblog

Gemma 3 QAT Models. Interesting release from Google, as a follow-up to Gemma 3 from last month:

To make Gemma 3 even more accessible, we are announcing new versions optimized with Quantization-Aware Training (QAT) that dramatically reduces memory requirements while maintaining high quality. This enables you to run powerful models like Gemma 3 27B locally on consumer-grade GPUs like the NVIDIA RTX 3090.

I wasn't previously aware of Quantization-Aware Training but it turns out to be quite an established pattern now, supported in both Tensorflow and PyTorch.

Google report model size drops from BF16 to int4 for the following models:

• Gemma 3 27B: 54GB to 14.1GB
• Gemma 3 12B: 24GB to 6.6GB
• Gemma 3 4B: 8GB to 2.6GB
• Gemma 3 1B: 2GB to 0.5GB

They partnered with Ollama, LM Studio, MLX (here's their collection) and llama.cpp for this release - I'd love to see more AI labs following their example.

The Ollama model version picker currently hides them behind "View all" option, so here are the direct links:

• gemma3:1b-it-qat - 1GB
• gemma3:4b-it-qat - 4GB
• gemma3:12b-it-qat - 8.9GB
• gemma3:27b-it-qat - 18GB

I fetched that largest model with:

ollama pull gemma3:27b-it-qat

And now I'm trying it out with llm-ollama:

llm -m gemma3:27b-it-qat "impress me with some physics"

I got a pretty great response!

Update: Having spent a while putting it through its paces via Open WebUI and Tailscale to access my laptop from my phone I think this may be my new favorite general-purpose local model. Ollama appears to use 22GB of RAM while the model is running, which leaves plenty on my 64GB machine for other applications.

I've also tried it via llm-mlx like this (downloading 16GB):

llm install llm-mlx
llm mlx download-model mlx-community/gemma-3-27b-it-qat-4bit
llm chat -m mlx-community/gemma-3-27b-it-qat-4bit

It feels a little faster with MLX and uses 15GB of memory according to Activity Monitor.

# 5:20 pm / google, ai, tailscale, generative-ai, local-llms, llms, llm, mlx, ollama, gemma, llm-release, lm-studio

Claude Code: Best practices for agentic coding (via) Extensive new documentation from Anthropic on how to get the best results out of their Claude Code CLI coding agent tool, which includes this fascinating tip:

We recommend using the word "think" to trigger extended thinking mode, which gives Claude additional computation time to evaluate alternatives more thoroughly. These specific phrases are mapped directly to increasing levels of thinking budget in the system: "think" < "think hard" < "think harder" < "ultrathink." Each level allocates progressively more thinking budget for Claude to use.

Apparently ultrathink is a magic word!

I was curious if this was a feature of the Claude model itself or Claude Code in particular. Claude Code isn't open source but you can view the obfuscated JavaScript for it, and make it a tiny bit less obfuscated by running it through Prettier. With Claude's help I used this recipe:

mkdir -p /tmp/claude-code-examine
cd /tmp/claude-code-examine
npm init -y
npm install @anthropic-ai/claude-code
cd node_modules/@anthropic-ai/claude-code
npx prettier --write cli.js

Then used ripgrep to search for "ultrathink":

rg ultrathink -C 30

And found this chunk of code:

let B = W.message.content.toLowerCase();
if (
  B.includes("think harder") ||
  B.includes("think intensely") ||
  B.includes("think longer") ||
  B.includes("think really hard") ||
  B.includes("think super hard") ||
  B.includes("think very hard") ||
  B.includes("ultrathink")
)
  return (
    l1("tengu_thinking", { tokenCount: 31999, messageId: Z, provider: G }),
    31999
  );
if (
  B.includes("think about it") ||
  B.includes("think a lot") ||
  B.includes("think deeply") ||
  B.includes("think hard") ||
  B.includes("think more") ||
  B.includes("megathink")
)
  return (
    l1("tengu_thinking", { tokenCount: 1e4, messageId: Z, provider: G }), 1e4
  );
if (B.includes("think"))
  return (
    l1("tengu_thinking", { tokenCount: 4000, messageId: Z, provider: G }),
    4000
  );

So yeah, it looks like "ultrathink" is a Claude Code feature - presumably that 31999 is a number that affects the token thinking budget, especially since "megathink" maps to 1e4 tokens (10,000) and just plain "think" maps to 4,000.

# 10:17 pm / cli, ai, generative-ai, llms, ai-assisted-programming, anthropic, claude, llm-reasoning, coding-agents, claude-code

llm-fragments-github 0.2. I upgraded my llm-fragments-github plugin to add a new fragment type called issue. It lets you pull the entire content of a GitHub issue thread into your prompt as a concatenated Markdown file.

(If you haven't seen fragments before I introduced them in Long context support in LLM 0.24 using fragments and template plugins.)

I used it just now to have Gemini 2.5 Pro provide feedback and attempt an implementation of a complex issue against my LLM project:

llm install llm-fragments-github
llm -f github:simonw/llm \
  -f issue:simonw/llm/938 \
  -m gemini-2.5-pro-exp-03-25 \
  --system 'muse on this issue, then propose a whole bunch of code to help implement it'

Here I'm loading the FULL content of the simonw/llm repo using that -f github:simonw/llm fragment (documented here), then loading all of the comments from issue 938 where I discuss quite a complex potential refactoring. I ask Gemini 2.5 Pro to "muse on this issue" and come up with some code.

This worked shockingly well. Here's the full response, which highlighted a few things I hadn't considered yet (such as the need to migrate old database records to the new tree hierarchy) and then spat out a whole bunch of code which looks like a solid start to the actual implementation work I need to do.

I ran this against Google's free Gemini 2.5 Preview, but if I'd used the paid model it would have cost me 202,680 input tokens, 10,460 output tokens and 1,859 thinking tokens for a total of 62.989 cents.

As a fun extra, the new issue: feature itself was written almost entirely by OpenAI o3, again using fragments. I ran this:

llm -m openai/o3 \
  -f https://raw.githubusercontent.com/simonw/llm-hacker-news/refs/heads/main/llm_hacker_news.py \
  -f https://raw.githubusercontent.com/simonw/tools/refs/heads/main/github-issue-to-markdown.html \
  -s 'Write a new fragments plugin in Python that registers issue:org/repo/123 which fetches that issue
      number from the specified github repo and uses the same markdown logic as the HTML page to turn that into a fragment'

Here I'm using the ability to pass a URL to -f and giving it the full source of my llm_hacker_news.py plugin (which shows how a fragment can load data from an API) plus the HTML source of my github-issue-to-markdown tool (which I wrote a few months ago with Claude). I effectively asked o3 to take that HTML/JavaScript tool and port it to Python to work with my fragments plugin mechanism.

o3 provided almost the exact implementation I needed, and even included support for a GITHUB_TOKEN environment variable without me thinking to ask for it. Total cost: 19.928 cents.

On a final note of curiosity I tried running this prompt against Gemma 3 27B QAT running on my Mac via MLX and llm-mlx:

llm install llm-mlx
llm mlx download-model mlx-community/gemma-3-27b-it-qat-4bit

llm -m mlx-community/gemma-3-27b-it-qat-4bit \
  -f https://raw.githubusercontent.com/simonw/llm-hacker-news/refs/heads/main/llm_hacker_news.py \
  -f https://raw.githubusercontent.com/simonw/tools/refs/heads/main/github-issue-to-markdown.html \
  -s 'Write a new fragments plugin in Python that registers issue:org/repo/123 which fetches that issue
      number from the specified github repo and uses the same markdown logic as the HTML page to turn that into a fragment'

That worked pretty well too. It turns out a 16GB local model file is powerful enough to write me an LLM plugin now!

# 2:01 pm / github, plugins, ai, generative-ai, local-llms, llms, ai-assisted-programming, llm, gemini, mlx, o3, long-context, gemma

Decentralizing Schemes. Tim Bray discusses the challenges faced by decentralized Mastodon in that shared URLs to posts don't take into account people accessing Mastodon via their own instances, which breaks replies/likes/shares etc unless you further copy and paste URLs around yourself.

Tim proposes that the answer is URIs: a registered fedi://mastodon.cloud/@timbray/109508984818551909 scheme could allow Fediverse-aware software to step in and handle those URIs, similar to how mailto: works.

Bluesky have registered at: already, and there's also a web+ap: prefix registered with the intent of covering ActivityPub, the protocol used by Mastodon.

# 6:48 pm / decentralisation, social-media, tim-bray, urls, mastodon, bluesky

OpenAI o3 and o4-mini System Card. I'm surprised to see a combined System Card for o3 and o4-mini in the same document - I'd expect to see these covered separately.

The opening paragraph calls out the most interesting new ability of these models (see also my notes here). Tool usage isn't new, but using tools in the chain of thought appears to result in some very significant improvements:

The models use tools in their chains of thought to augment their capabilities; for example, cropping or transforming images, searching the web, or using Python to analyze data during their thought process.

Section 3.3 on hallucinations has been gaining a lot of attention. Emphasis mine:

We tested OpenAI o3 and o4-mini against PersonQA, an evaluation that aims to elicit hallucinations. PersonQA is a dataset of questions and publicly available facts that measures the model's accuracy on attempted answers.

We consider two metrics: accuracy (did the model answer the question correctly) and hallucination rate (checking how often the model hallucinated).

The o4-mini model underperforms o1 and o3 on our PersonQA evaluation. This is expected, as smaller models have less world knowledge and tend to hallucinate more. However, we also observed some performance differences comparing o1 and o3. Specifically, o3 tends to make more claims overall, leading to more accurate claims as well as more inaccurate/hallucinated claims. More research is needed to understand the cause of this result.

Table 4: PersonQA evaluation

Metric	o3	o4-mini	o1
accuracy (higher is better)	0.59	0.36	0.47
hallucination rate (lower is better)	0.33	0.48	0.16

The benchmark score on OpenAI's internal PersonQA benchmark (as far as I can tell no further details of that evaluation have been shared) going from 0.16 for o1 to 0.33 for o3 is interesting, but I don't know if it it's interesting enough to produce dozens of headlines along the lines of "OpenAI's o3 and o4-mini hallucinate way higher than previous models".

The paper also talks at some length about "sandbagging". I’d previously encountered sandbagging defined as meaning “where models are more likely to endorse common misconceptions when their user appears to be less educated”. The o3/o4-mini system card uses a different definition: “the model concealing its full capabilities in order to better achieve some goal” - and links to the recent Anthropic paper Automated Researchers Can Subtly Sandbag.

As far as I can tell this definition relates to the American English use of “sandbagging” to mean “to hide the truth about oneself so as to gain an advantage over another” - as practiced by poker or pool sharks.

(Wouldn't it be nice if we could have just one piece of AI terminology that didn't attract multiple competing definitions?)

o3 and o4-mini both showed some limited capability to sandbag - to attempt to hide their true capabilities in safety testing scenarios that weren't fully described. This relates to the idea of "scheming", which I wrote about with respect to the GPT-4o model card last year.

# 7:13 pm / ai, openai, generative-ai, llms, o3, ai-ethics, hallucinations

Working Through the Fear of Being Seen (via) Heartfelt piece by Ashley Willis about the challenge of overcoming self-doubt in publishing online:

Part of that is knowing who might read it. A lot of the folks who follow me are smart, opinionated, and not always generous. Some are friends. Some are people I’ve looked up to. And some are just really loud on the internet. I saw someone the other day drag a certain writing style. That kind of judgment makes me want to shrink back and say, never mind.

Try to avoid being somebody who discourages others from sharing their thoughts.

# 6:40 am / blogging

A5 (via) A5 is a new "global, equal-area, millimeter-accurate geospatial index" by Felix Palmer:

It is the pentagonal equivalent of other DGGSs, like S2 or H3, but with higher accuracy and lower distortion.

Effectively it's a way of dividing the entire world into pentagons where each one covers the same physical area (to within a 2% threshold) - like Uber's H3 but a bit weirder and more fun. An A5 reference implementation written in TypeScript is available on GitHub.

This interactive demo helps show how it works:

Why pentagons? Here's what the A5 docs say:

A5 is unique in that it uses a pentagonal tiling of a dodecahedron. [...] The benefit of choosing a dodecahedron is that it is the platonic solid with the lowest vertex curvature, and by this measure it is the most spherical of all the platonic solids. This is key for minimizing cell distortion as the process of projecting a platonic solid onto a sphere involves warping the cell geometry to force the vertex curvature to approach zero. Thus, the lower the original vertex curvature, the less distortion will be introduced by the projection.

I had to look up platonic solids on Wikipedia. There are only five: Tetrahedron, Cube, Octahedron, Dodecahedron and Icosahedron and they can be made using squares, triangles or (in the case of the Dodecahedron) pentagons, making the pentagon the most circle-like option.

# 2:37 pm / geospatial

Abusing DuckDB-WASM by making SQL draw 3D graphics (Sort Of) (via) Brilliant hack by Patrick Trainer who got an ASCII-art Doom clone running in the browser using convoluted SQL queries running against the WebAssembly build of DuckDB. Here’s the live demo, and the code on GitHub.

The SQL is so much fun. Here’s a snippet that implements ray tracing as part of a SQL view:

CREATE OR REPLACE VIEW render_3d_frame AS
WITH RECURSIVE
    -- ...
    rays AS (
        SELECT 
            c.col, 
            (p.dir - s.fov/2.0 + s.fov * (c.col*1.0 / (s.view_w - 1))) AS angle 
        FROM cols c, s, p
    ),
    raytrace(col, step_count, fx, fy, angle) AS (
        SELECT 
            r.col, 
            1, 
            p.x + COS(r.angle)*s.step, 
            p.y + SIN(r.angle)*s.step, 
            r.angle 
        FROM rays r, p, s 
        UNION ALL 
        SELECT 
            rt.col, 
            rt.step_count + 1, 
            rt.fx + COS(rt.angle)*s.step, 
            rt.fy + SIN(rt.angle)*s.step, 
            rt.angle 
        FROM raytrace rt, s 
        WHERE rt.step_count < s.max_steps 
          AND NOT EXISTS (
              SELECT 1 
              FROM map m 
              WHERE m.x = CAST(rt.fx AS INT) 
                AND m.y = CAST(rt.fy AS INT) 
                AND m.tile = '#'
          )
    ),
    -- ...

# 4:29 pm / sql, webassembly, duckdb

ClickHouse gets lazier (and faster): Introducing lazy materialization (via) Tom Schreiber describe's the latest optimization in ClickHouse, and in the process explores a whole bunch of interesting characteristics of columnar datastores generally.

As I understand it, the new "lazy materialization" feature means that if you run a query like this:

select id, big_col1, big_col2
from big_table order by rand() limit 5

Those big_col1 and big_col2 columns won't be read from disk for every record, just for the five that are returned. This can dramatically improve the performance of queries against huge tables - for one example query ClickHouse report a drop from "219 seconds to just 139 milliseconds—with 40× less data read and 300× lower memory usage."

I'm linking to this mainly because the article itself is such a detailed discussion of columnar data patterns in general. It caused me to update my intuition for how queries against large tables can work on modern hardware. This query for example:

SELECT helpful_votes
FROM amazon.amazon_reviews
ORDER BY helpful_votes DESC
LIMIT 3;

Can run in 70ms against a 150 million row, 70GB table - because in a columnar database you only need to read that helpful_votes integer column which adds up to just 600MB of data, and sorting 150 million integers on a decent machine takes no time at all.

# 5:05 pm / databases, sql, clickhouse