1,030 items tagged “ai”
2024
OpenAI DevDay: Let’s build developer tools, not digital God
I had a fun time live blogging OpenAI DevDay yesterday—I’ve now shared notes about the live blogging system I threw other in a hurry on the day (with assistance from Claude and GPT-4o). Now that the smoke has settled a little, here are my impressions from the event.
[... 2,090 words]Ethical Applications of AI to Public Sector Problems. Jacob Kaplan-Moss developed this model a few years ago (before the generative AI rush) while working with public-sector startups and is publishing it now. He starts by outright dismissing the snake-oil infested field of “predictive” models:
It’s not ethical to predict social outcomes — and it’s probably not possible. Nearly everyone claiming to be able to do this is lying: their algorithms do not, in fact, make predictions that are any better than guesswork. […] Organizations acting in the public good should avoid this area like the plague, and call bullshit on anyone making claims of an ability to predict social behavior.
Jacob then differentiates assistive AI and automated AI. Assistive AI helps human operators process and consume information, while leaving the human to take action on it. Automated AI acts upon that information without human oversight.
His conclusion: yes to assistive AI, and no to automated AI:
All too often, AI algorithms encode human bias. And in the public sector, failure carries real life or death consequences. In the private sector, companies can decide that a certain failure rate is OK and let the algorithm do its thing. But when citizens interact with their governments, they have an expectation of fairness, which, because AI judgement will always be available, it cannot offer.
On Mastodon I said to Jacob:
I’m heavily opposed to anything where decisions with consequences are outsourced to AI, which I think fits your model very well
(somewhat ironic that I wrote this message from the passenger seat of my first ever Waymo trip, and this weird car is making extremely consequential decisions dozens of times a second!)
Which sparked an interesting conversation about why life-or-death decisions made by self-driving cars feel different from decisions about social services. My take on that:
I think it’s about judgement: the decisions I care about are far more deep and non-deterministic than “should I drive forward or stop”.
Where there’s moral ambiguity, I want a human to own the decision both so there’s a chance for empathy, and also for someone to own the accountability for the choice.
That idea of ownership and accountability for decision making feels critical to me. A giant black box of matrix multiplication cannot take accountability for “decisions” that it makes.
Building an automatically updating live blog in Django. Here's an extended write-up of how I implemented the live blog feature I used for my coverage of OpenAI DevDay yesterday. I built the first version using Claude while waiting for the keynote to start, then upgraded it during the lunch break with the help of GPT-4o to add sort options and incremental fetching of new updates.
OpenAI DevDay 2024 live blog
I’m at OpenAI DevDay in San Francisco, and I’m trying something new: a live blog, where this entry will be updated with new notes during the event.
[... 68 words]Whisper large-v3-turbo model. It’s OpenAI DevDay today. Last year they released a whole stack of new features, including GPT-4 vision and GPTs and their text-to-speech API, so I’m intrigued to see what they release today (I’ll be at the San Francisco event).
Looks like they got an early start on the releases, with the first new Whisper model since November 2023.
Whisper Turbo is a new speech-to-text model that fits the continued trend of distilled models getting smaller and faster while maintaining the same quality as larger models.
large-v3-turbo is 809M parameters - slightly larger than the 769M medium but significantly smaller than the 1550M large. OpenAI claim its 8x faster than large and requires 6GB of VRAM compared to 10GB for the larger model.
The model file is a 1.6GB download. OpenAI continue to make Whisper (both code and model weights) available under the MIT license.
It’s already supported in both Hugging Face transformers - live demo here - and in mlx-whisper on Apple Silicon, via Awni Hannun:
import mlx_whisper
print(mlx_whisper.transcribe(
"path/to/audio",
path_or_hf_repo="mlx-community/whisper-turbo"
)["text"])
Awni reports:
Transcribes 12 minutes in 14 seconds on an M2 Ultra (~50X faster than real time).
I listened to the whole 15-minute podcast this morning. It was, indeed, surprisingly effective. It remains somewhere in the uncanny valley, but not at all in a creepy way. Just more in a “this is a bit vapid and phony” way. [...] But ultimately the conversation has all the flavor of a bowl of unseasoned white rice.
Conflating Overture Places Using DuckDB, Ollama, Embeddings, and More.
Drew Breunig's detailed tutorial on "conflation" - combining different geospatial data sources by de-duplicating address strings such as RESTAURANT LOS ARCOS,3359 FOOTHILL BLVD,OAKLAND,94601 and LOS ARCOS TAQUERIA,3359 FOOTHILL BLVD,OAKLAND,94601.
Drew uses an entirely offline stack based around Python, DuckDB and Ollama and finds that a combination of H3 geospatial tiles and mxbai-embed-large embeddings (though other embedding models should work equally well) gets really good results.
llama-3.2-webgpu (via) Llama 3.2 1B is a really interesting models, given its 128,000 token input and its tiny size (barely more than a GB).
This page loads a 1.24GB q4f16 ONNX build of the Llama-3.2-1B-Instruct model and runs it with a React-powered chat interface directly in the browser, using Transformers.js and WebGPU. Source code for the demo is here.
It worked for me just now in Chrome; in Firefox and Safari I got a “WebGPU is not supported by this browser” error message.
NotebookLM’s automatically generated podcasts are surprisingly effective
Audio Overview is a fun new feature of Google’s NotebookLM which is getting a lot of attention right now. It generates a one-off custom podcast against content you provide, where two AI hosts start up a “deep dive” discussion about the collected content. These last around ten minutes and are very podcast, with an astonishingly convincing audio back-and-forth conversation.
[... 1,489 words]mlx-vlm (via) The MLX ecosystem of libraries for running machine learning models on Apple Silicon continues to expand. Prince Canuma is actively developing this library for running vision models such as Qwen-2 VL and Pixtral and LLaVA using Python running on a Mac.
I used uv to run it against this image with this shell one-liner:
uv run --with mlx-vlm \
python -m mlx_vlm.generate \
--model Qwen/Qwen2-VL-2B-Instruct \
--max-tokens 1000 \
--temp 0.0 \
--image https://static.simonwillison.net/static/2024/django-roadmap.png \
--prompt "Describe image in detail, include all text"
The --image option works equally well with a URL or a path to a local file on disk.
This first downloaded 4.1GB to my ~/.cache/huggingface/hub/models--Qwen--Qwen2-VL-2B-Instruct folder and then output this result, which starts:
The image is a horizontal timeline chart that represents the release dates of various software versions. The timeline is divided into years from 2023 to 2029, with each year represented by a vertical line. The chart includes a legend at the bottom, which distinguishes between different types of software versions. [...]
In the future, we won't need programmers; just people who can describe to a computer precisely what they want it to do.
OpenAI’s revenue in August more than tripled from a year ago, according to the documents, and about 350 million people — up from around 100 million in March — used its services each month as of June. […]
Roughly 10 million ChatGPT users pay the company a $20 monthly fee, according to the documents. OpenAI expects to raise that price by $2 by the end of the year, and will aggressively raise it to $44 over the next five years, the documents said.
I think individual creators or publishers tend to overestimate the value of their specific content in the grand scheme of [AI training]. […]
We pay for content when it’s valuable to people. We’re just not going to pay for content when it’s not valuable to people. I think that you’ll probably see a similar dynamic with AI, which my guess is that there are going to be certain partnerships that get made when content is really important and valuable. I’d guess that there are probably a lot of people who have a concern about the feel of it, like you’re saying. But then, when push comes to shove, if they demanded that we don’t use their content, then we just wouldn’t use their content. It’s not like that’s going to change the outcome of this stuff that much.
Llama 3.2. In further evidence that AI labs are terrible at naming things, Llama 3.2 is a huge upgrade to the Llama 3 series - they've released their first multi-modal vision models!
Today, we’re releasing Llama 3.2, which includes small and medium-sized vision LLMs (11B and 90B), and lightweight, text-only models (1B and 3B) that fit onto edge and mobile devices, including pre-trained and instruction-tuned versions.
The 1B and 3B text-only models are exciting too, with a 128,000 token context length and optimized for edge devices (Qualcomm and MediaTek hardware get called out specifically).
Meta partnered directly with Ollama to help with distribution, here's the Ollama blog post. They only support the two smaller text-only models at the moment - this command will get the 3B model (2GB):
ollama run llama3.2
And for the 1B model (a 1.3GB download):
ollama run llama3.2:1b
I had to first upgrade my Ollama by clicking on the icon in my macOS task tray and selecting "Restart to update".
The two vision models are coming to Ollama "very soon".
Once you have fetched the Ollama model you can access it from my LLM command-line tool like this:
pipx install llm
llm install llm-ollama
llm chat -m llama3.2:1b
I tried running my djp codebase through that tiny 1B model just now and got a surprisingly good result - by no means comprehensive, but way better than I would ever expect from a model of that size:
files-to-prompt **/*.py -c | llm -m llama3.2:1b --system 'describe this code'
Here's a portion of the output:
The first section defines several test functions using the
@djp.hookimpldecorator from the djp library. These hook implementations allow you to intercept and manipulate Django's behavior.
test_middleware_order: This function checks that the middleware order is correct by comparing theMIDDLEWAREsetting with a predefined list.test_middleware: This function tests various aspects of middleware:- It retrieves the response from the URL
/from-plugin/using theClientobject, which simulates a request to this view.- It checks that certain values are present in the response:
X-DJP-Middleware-AfterX-DJP-MiddlewareX-DJP-Middleware-Before[...]
I found the GGUF file that had been downloaded by Ollama in my ~/.ollama/models/blobs directory. The following command let me run that model directly in LLM using the llm-gguf plugin:
llm install llm-gguf
llm gguf register-model ~/.ollama/models/blobs/sha256-74701a8c35f6c8d9a4b91f3f3497643001d63e0c7a84e085bed452548fa88d45 -a llama321b
llm chat -m llama321b
Meta themselves claim impressive performance against other existing models:
Our evaluation suggests that the Llama 3.2 vision models are competitive with leading foundation models, Claude 3 Haiku and GPT4o-mini on image recognition and a range of visual understanding tasks. The 3B model outperforms the Gemma 2 2.6B and Phi 3.5-mini models on tasks such as following instructions, summarization, prompt rewriting, and tool-use, while the 1B is competitive with Gemma.
Here's the Llama 3.2 collection on Hugging Face. You need to accept the new Llama 3.2 Community License Agreement there in order to download those models.
You can try the four new models out via the Chatbot Arena - navigate to "Direct Chat" there and select them from the dropdown menu. You can upload images directly to the chat there to try out the vision features.
Solving a bug with o1-preview, files-to-prompt and LLM.
I added a new feature to DJP this morning: you can now have plugins specify their middleware in terms of how it should be positioned relative to other middleware - inserted directly before or directly after django.middleware.common.CommonMiddleware for example.
At one point I got stuck with a weird test failure, and after ten minutes of head scratching I decided to pipe the entire thing into OpenAI's o1-preview to see if it could spot the problem. I used files-to-prompt to gather the code and LLM to run the prompt:
files-to-prompt **/*.py -c | llm -m o1-preview "
The middleware test is failing showing all of these - why is MiddlewareAfter repeated so many times?
['MiddlewareAfter', 'Middleware3', 'MiddlewareAfter', 'Middleware5', 'MiddlewareAfter', 'Middleware3', 'MiddlewareAfter', 'Middleware2', 'MiddlewareAfter', 'Middleware3', 'MiddlewareAfter', 'Middleware5', 'MiddlewareAfter', 'Middleware3', 'MiddlewareAfter', 'Middleware4', 'MiddlewareAfter', 'Middleware3', 'MiddlewareAfter', 'Middleware5', 'MiddlewareAfter', 'Middleware3', 'MiddlewareAfter', 'Middleware2', 'MiddlewareAfter', 'Middleware3', 'MiddlewareAfter', 'Middleware5', 'MiddlewareAfter', 'Middleware3', 'MiddlewareAfter', 'Middleware', 'MiddlewareBefore']"The model whirled away for a few seconds and spat out an explanation of the problem - one of my middleware classes was accidentally calling self.get_response(request) in two different places.
I did enjoy how o1 attempted to reference the relevant Django documentation and then half-repeated, half-hallucinated a quote from it:
This took 2,538 input tokens and 4,354 output tokens - by my calculations at $15/million input and $60/million output that prompt cost just under 30 cents.
The Pragmatic Engineer Podcast: AI tools for software engineers, but without the hype – with Simon Willison. Gergely Orosz has a brand new podcast, and I was the guest for the first episode. We covered a bunch of ground, but my favorite topic was an exploration of the (very legitimate) reasons that many engineers are resistant to taking advantage of AI-assisted programming tools.
Updated production-ready Gemini models.
Two new models from Google Gemini today: gemini-1.5-pro-002 and gemini-1.5-flash-002. Their -latest aliases will update to these new models in "the next few days", and new -001 suffixes can be used to stick with the older models. The new models benchmark slightly better in various ways and should respond faster.
Flash continues to have a 1,048,576 input token and 8,192 output token limit. Pro is 2,097,152 input tokens.
Google also announced a significant price reduction for Pro, effective on the 1st of October. Inputs less than 128,000 tokens drop from $3.50/million to $1.25/million (above 128,000 tokens it's dropping from $7 to $5) and output costs drop from $10.50/million to $2.50/million ($21 down to $10 for the >128,000 case).
For comparison, GPT-4o is currently $5/m input and $15/m output and Claude 3.5 Sonnet is $3/m input and $15/m output. Gemini 1.5 Pro was already the cheapest of the frontier models and now it's even cheaper.
Correction: I missed gpt-4o-2024-08-06 which is listed later on the OpenAI pricing page and priced at $2.50/m input and $10/m output. So the new Gemini 1.5 Pro prices are undercutting that.
Gemini has always offered finely grained safety filters - it sounds like those are now turned down to minimum by default, which is a welcome change:
For the models released today, the filters will not be applied by default so that developers can determine the configuration best suited for their use case.
Also interesting: they've tweaked the expected length of default responses:
For use cases like summarization, question answering, and extraction, the default output length of the updated models is ~5-20% shorter than previous models.
XKCD 1425 (Tasks) turns ten years old today (via) One of the all-time great XKCDs. It's amazing that "check whether the photo is of a bird" has gone from PhD-level to trivially easy to solve (with a vision LLM, or CLIP, or ResNet+ImageNet among others).
The key idea still very much stands though. Understanding the difference between easy and hard challenges in software development continues to require an enormous depth of experience.
I'd argue that LLMs have made this even worse.
Understanding what kind of tasks LLMs can and cannot reliably solve remains incredibly difficult and unintuitive. They're computer systems that are terrible at maths and that can't reliably lookup facts!
On top of that, the rise of AI-assisted programming tools means more people than ever are beginning to create their own custom software.
These brand new AI-assisted proto-programmers are having a crash course in this easy-v.s.-hard problem.
I saw someone recently complaining that they couldn't build a Claude Artifact that could analyze images, even though they knew Claude itself could do that. Understanding why that's not possible involves understanding how the CSP headers that are used to serve Artifacts prevent the generated code from making its own API calls out to an LLM!
Whether you think coding with AI works today or not doesn’t really matter.
But if you think functional AI helping to code will make humans dumber or isn’t real programming just consider that’s been the argument against every generation of programming tools going back to Fortran.
Markdown and Math Live Renderer.
Another of my tiny Claude-assisted JavaScript tools. This one lets you enter Markdown with embedded mathematical expressions (like $ax^2 + bx + c = 0$) and live renders those on the page, with an HTML version using MathML that you can export through copy and paste.
Here's the Claude transcript. I started by asking:
Are there any client side JavaScript markdown libraries that can also handle inline math and render it?
Claude gave me several options including the combination of Marked and KaTeX, so I followed up by asking:
Build an artifact that demonstrates Marked plus KaTeX - it should include a text area I can enter markdown in (repopulated with a good example) and live update the rendered version below. No react.
Which gave me this artifact, instantly demonstrating that what I wanted to do was possible.
I iterated on it a tiny bit to get to the final version, mainly to add that HTML export and a Copy button. The final source code is here.
YouTube Thumbnail Viewer.
I wanted to find the best quality thumbnail image for a YouTube video, so I could use it as a social media card. I know from past experience that GPT-4 has memorized the various URL patterns for img.youtube.com, so I asked it to guess the URL for my specific video.
This piqued my interest as to what the other patterns were, so I got it to spit those out too. Then, to save myself from needing to look those up again in the future, I asked it to build me a little HTML and JavaScript tool for turning a YouTube video URL into a set of visible thumbnails.
I iterated on the code a bit more after pasting it into Claude and ended up with this, now hosted in my tools collection.
Notes on using LLMs for code
I was recently the guest on TWIML—the This Week in Machine Learning & AI podcast. Our episode is titled Supercharging Developer Productivity with ChatGPT and Claude with Simon Willison, and the focus of the conversation was the ways in which I use LLM tools in my day-to-day work as a software developer and product engineer.
[... 861 words]Introducing Contextual Retrieval (via) Here's an interesting new embedding/RAG technique, described by Anthropic but it should work for any embedding model against any other LLM.
One of the big challenges in implementing semantic search against vector embeddings - often used as part of a RAG system - is creating "chunks" of documents that are most likely to semantically match queries from users.
Anthropic provide this solid example where semantic chunks might let you down:
Imagine you had a collection of financial information (say, U.S. SEC filings) embedded in your knowledge base, and you received the following question: "What was the revenue growth for ACME Corp in Q2 2023?"
A relevant chunk might contain the text: "The company's revenue grew by 3% over the previous quarter." However, this chunk on its own doesn't specify which company it's referring to or the relevant time period, making it difficult to retrieve the right information or use the information effectively.
Their proposed solution is to take each chunk at indexing time and expand it using an LLM - so the above sentence would become this instead:
This chunk is from an SEC filing on ACME corp's performance in Q2 2023; the previous quarter's revenue was $314 million. The company's revenue grew by 3% over the previous quarter.
This chunk was created by Claude 3 Haiku (their least expensive model) using the following prompt template:
<document>
{{WHOLE_DOCUMENT}}
</document>
Here is the chunk we want to situate within the whole document
<chunk>
{{CHUNK_CONTENT}}
</chunk>
Please give a short succinct context to situate this chunk within the overall document for the purposes of improving search retrieval of the chunk. Answer only with the succinct context and nothing else.
Here's the really clever bit: running the above prompt for every chunk in a document could get really expensive thanks to the inclusion of the entire document in each prompt. Claude added context caching last month, which allows you to pay around 1/10th of the cost for tokens cached up to your specified beakpoint.
By Anthropic's calculations:
Assuming 800 token chunks, 8k token documents, 50 token context instructions, and 100 tokens of context per chunk, the one-time cost to generate contextualized chunks is $1.02 per million document tokens.
Anthropic provide a detailed notebook demonstrating an implementation of this pattern. Their eventual solution combines cosine similarity and BM25 indexing, uses embeddings from Voyage AI and adds a reranking step powered by Cohere.
The notebook also includes an evaluation set using JSONL - here's that evaluation data in Datasette Lite.
Moshi (via) Moshi is "a speech-text foundation model and full-duplex spoken dialogue framework". It's effectively a text-to-text model - like an LLM but you input audio directly to it and it replies with its own audio.
It's fun to play around with, but it's not particularly useful in comparison to other pure text models: I tried to talk to it about California Brown Pelicans and it gave me some very basic hallucinated thoughts about California Condors instead.
It's very easy to run locally, at least on a Mac (and likely on other systems too). I used uv and got the 8 bit quantized version running as a local web server using this one-liner:
uv run --with moshi_mlx python -m moshi_mlx.local_web -q 8
That downloads ~8.17G of model to a folder in ~/.cache/huggingface/hub/ - or you can use -q 4 and get a 4.81G version instead (albeit even lower quality).
The problem that you face is that it's relatively easy to take a model and make it look like it's aligned. You ask GPT-4, “how do I end all of humans?” And the model says, “I can't possibly help you with that”. But there are a million and one ways to take the exact same question - pick your favorite - and you can make the model still answer the question even though initially it would have refused. And the question this reminds me a lot of coming from adversarial machine learning. We have a very simple objective: Classify the image correctly according to the original label. And yet, despite the fact that it was essentially trivial to find all of the bugs in principle, the community had a very hard time coming up with actually effective defenses. We wrote like over 9,000 papers in ten years, and have made very very very limited progress on this one small problem. You all have a harder problem and maybe less time.
Supercharging Developer Productivity with ChatGPT and Claude with Simon Willison (via) I'm the guest for the latest episode of the TWIML AI podcast - This Week in Machine Learning & AI, hosted by Sam Charrington.
We mainly talked about how I use LLM tooling for my own work - Claude, ChatGPT, Code Interpreter, Claude Artifacts, LLM and GitHub Copilot - plus a bit about my experiments with local models.
o1 prompting is alien to me. Its thinking, gloriously effective at times, is also dreamlike and unamenable to advice.
Just say what you want and pray. Any notes on “how” will be followed with the diligence of a brilliant intern on ketamine.
Speed matters (via) Jamie Brandon in 2021, talking about the importance of optimizing for the speed at which you can work as a developer:
Being 10x faster also changes the kinds of projects that are worth doing.
Last year I spent something like 100 hours writing a text editor. […] If I was 10x slower it would have been 20-50 weeks. Suddenly that doesn't seem like such a good deal any more - what a waste of a year!
It’s not just about speed of writing code:
When I think about speed I think about the whole process - researching, planning, designing, arguing, coding, testing, debugging, documenting etc.
Often when I try to convince someone to get faster at one of those steps, they'll argue that the others are more important so it's not worthwhile trying to be faster. Eg choosing the right idea is more important than coding the wrong idea really quickly.
But that's totally conditional on the speed of everything else! If you could code 10x as fast then you could try out 10 different ideas in the time it would previously have taken to try out 1 idea. Or you could just try out 1 idea, but have 90% of your previous coding time available as extra idea time.
Jamie’s model here helps explain the effect I described in AI-enhanced development makes me more ambitious with my projects. Prompting an LLM to write portions of my code for me gives me that 5-10x boost in the time I spend typing code into a computer, which has a big effect on my ambitions despite being only about 10% of the activities I perform relevant to building software.
I also increasingly lean on LLMs as assistants in the research phase - exploring library options, building experimental prototypes - and for activities like writing tests and even a little bit of documentation.
[… OpenAI’s o1] could work its way to a correct (and well-written) solution if provided a lot of hints and prodding, but did not generate the key conceptual ideas on its own, and did make some non-trivial mistakes. The experience seemed roughly on par with trying to advise a mediocre, but not completely incompetent, graduate student. However, this was an improvement over previous models, whose capability was closer to an actually incompetent graduate student.
It's a bit sad and confusing that LLMs ("Large Language Models") have little to do with language; It's just historical. They are highly general purpose technology for statistical modeling of token streams. A better name would be Autoregressive Transformers or something.
They don't care if the tokens happen to represent little text chunks. It could just as well be little image patches, audio chunks, action choices, molecules, or whatever. If you can reduce your problem to that of modeling token streams (for any arbitrary vocabulary of some set of discrete tokens), you can "throw an LLM at it".