Leaked Internal Google Document Claims Open Source AI Will Outcompete Google And OpenAI

May 5, 2023 by Maya Posch 37 Comments

In the world of large language models (LLM), the focus has for the longest time been on proprietary technologies from companies such as OpenAI (GPT-3 & 4, ChatGPT, etc.) as well as increasingly everyone from Google to Meta and Microsoft. What’s remained underexposed in this whole discussion about which LLM will do more things better are the efforts by hobbyists, unaffiliated researchers and everyone else you may find in Open Source LLM projects. According to a leaked document from a researcher at Google (anonymous, but apparently verified), Google is very worried that Open Source LLMs will wipe the floor with both Google’s and OpenAI’s efforts.

According to the document, after the open source community got their hands on the leaked LLaMA foundation model, motivated and highly knowledgeable individuals set to work to take a fairly basic model to new levels where it could begin to compete with the offerings by OpenAI and Google. Major innovations are the scaling issues, allowing these LLMs to work on far less powerful systems (like a laptop or even smartphone).

An important factor here is Low-Rank adaptation (LoRa), which massively cuts down the effort and resources required to train a model. Ultimately, as this document phrases it, Google and in extension OpenAI do not have a ‘secret sauce’ that makes their approaches better than anything the wider community can come up with. Noted is also that essentially Meta has won out here by having their LLM leak, as it has meant that the OSS community has been improving on the Meta foundations, allowing Meta to benefit from those improvements in their products.

The dire prediction is thus that in the end the proprietary LLMs by Google, OpenAI and others will cease to be relevant, as the open source community will have steamrolled them into fine, digital dust. Whether this will indeed work out this way remains to be seen, but things are not looking up for proprietary LLMs.

(Thanks to [Mike Szczys] for the tip)

ChatGPT Makes A 3D Model: The Secret Ingredient? Much Patience

May 4, 2023 by Donald Papp 44 Comments

ChatGPT is an AI large language model (LLM) which specializes in conversation. While using it, [Gil Meiri] discovered that one way to create models in FreeCAD is with Python scripting, and ChatGPT could be encouraged to create a 3D model of a plane in FreeCAD by expressing the model as a script. The result is just a basic plane shape, and it certainly took a lot of guidance on [Gil]’s part to make it happen, but it’s not bad for a tool that can’t see what it is doing.

The first step was getting ChatGPT to create code for a 10 mm cube, and plug that in FreeCAD to see the results. After that basic workflow was shown to work, [Gil] asked it to create a simple airplane shape. The resulting code had objects for wing, fuselage, and tail, but that’s about all that could be said because the result was almost — but not quite — completely unlike a plane. Not an encouraging start, but at least the basic building blocks were there. Continue reading “ChatGPT Makes A 3D Model: The Secret Ingredient? Much Patience” →

Chatting With Local AI Moves Directly In-Browser, Thanks To Web LLM

April 24, 2023 by Donald Papp 7 Comments

Large Language Models (LLM) are at the heart of natural-language AI tools like ChatGPT, and Web LLM shows it is now possible to run an LLM directly in a browser. Just to be clear, this is not a browser front end talking via API to some server-side application. This is a client-side LLM running entirely in the browser.

The ability to run an LLM (natural language AI) directly in-browser means more ways to implement local AI while enjoying GPU acceleration via WebGPU.

Running an AI system like an LLM locally usually leverages the computational abilities of a graphics card (GPU) to accelerate performance. This is true when running an image-generating AI system like Stable Diffusion, and it’s also true when implementing a local copy of an LLM like Vicuna (which happens to be the model implemented by Web LLM.) The thing that made Web LLM possible is WebGPU, whose release we covered just last month.

WebGPU provides a way for an in-browser application to talk to a local GPU directly, and it sure didn’t take long for someone to get the idea of using that to get a local LLM to run entirely within the browser, complete with GPU acceleration. This approach isn’t just limited to language models, either. The same method has been applied to successfully create Web Stable Diffusion as well.

It’s a fascinating (and fast) development that opens up new possibilities and, hopefully, gives people some new ideas. Check out Web LLM’s GitHub repository for a closer look, as well as access to an online demo.

Peering Down Into Talking Ant Hill

April 19, 2023 by Matthew Carlson 14 Comments

Watching an anthill brings an air of fascination. Thousands of ants are moving about and communicating with other ants as they work towards a goal as a collective whole. For us humans, we project a complex inner world for each of these tiny creatures to drive the narrative. But what if we could peer down into a miniature world and the ants spoke English? (PDF whitepaper)

Researchers at the University of Stanford and Google Research have released a paper about simulating human behavior using multiple Large Language Models (LMM). The simulation has a few dozen agents that can move across the small town, do errands, and communicate with each other. Each agent has a short description to help provide context to the LLM. In addition, they have memories of objects, other agents, and observations that they can retrieve, which allows them to create a plan for their day. The memory is a time-stamped text stream that the agent reflects on, deciding what is important. Additionally, the LLM can replan and figure out what it wants to do.

The question is, does the simulation seem life-like? One fascinating example is the paper’s authors created one agent (Isabella) intending to have a Valentine’s Day party. No other information is included. But several agents arrive at the character’s house later in the day to party. Isabella invited friends, and those agents asked some people.

A demo using recorded data from an earlier demo is web-accessible. However, it doesn’t showcase the powers that a user can exert on the world when running live. Thoughts and suggestions can be issued to an agent to steer their actions. However, you can pause the simulation to view the conversations between agents. Overall, it is incredible how life-like the simulation can be. The language of the conversation is quite formal, and running the simulation burns significant amounts of computing power. Perhaps there can be a subconscious where certain behaviors or observations can be coded in the agent instead of querying the LLM for every little thing (which sort of sounds like what people do).

There’s been an exciting trend of combining LLMs with a form of backing store, like combining Wolfram Alpha with chatGPT. Thanks [Abe] for sending this one in!

Tired Of Web Scraping? Make The AI Do It

April 9, 2023 by Donald Papp 19 Comments

[James Turk] has a novel approach to the problem of scraping web content in a structured way without needing to write the kind of page-specific code web scrapers usually have to deal with. How? Just enlist the help of a natural language AI. Scrapeghost relies on OpenAI’s GPT API to parse a web page’s content, pull out and classify any salient bits, and format it in a useful way.

What makes Scrapeghost different is how data gets organized. For example, when instantiating scrapeghost one defines the data one wishes to extract. For example:

from scrapeghost import SchemaScraper
scrape_legislators = SchemaScraper(
schema={
"name": "string",
"url": "url",
"district": "string",
"party": "string",
"photo_url": "url",
"offices": [{"name": "string", "address": "string", "phone": "string"}],
}
)

The kicker is that this format is entirely up to you! The GPT models are very, very good at processing natural language, and scrapeghost uses GPT to process the scraped data and find (using the example above) whatever looks like a name, district, party, photo, and office address and format it exactly as requested.

It’s an experimental tool and you’ll need an API key from OpenAI to use it, but it has useful features and is certainly a novel approach. There’s a tutorial and even a command-line interface, so check it out.

Why LLaMa Is A Big Deal

March 22, 2023 by Matthew Carlson 27 Comments

You might have heard about LLaMa or maybe you haven’t. Either way, what’s the big deal? It’s just some AI thing. In a nutshell, LLaMa is important because it allows you to run large language models (LLM) like GPT-3 on commodity hardware. In many ways, this is a bit like Stable Diffusion, which similarly allowed normal folks to run image generation models on their own hardware with access to the underlying source code. We’ve discussed why Stable Diffusion matters and even talked about how it works.

LLaMa is a transformer language model from Facebook/Meta research, which is a collection of large models from 7 billion to 65 billion parameters trained on publicly available datasets. Their research paper showed that the 13B version outperformed GPT-3 in most benchmarks and LLama-65B is right up there with the best of them. LLaMa was unique as inference could be run on a single GPU due to some optimizations made to the transformer itself and the model being about 10x smaller. While Meta recommended that users have at least 10 GB of VRAM to run inference on the larger models, that’s a huge step from the 80 GB A100 cards that often run these models.

While this was an important step forward for the research community, it became a huge one for the hacker community when [Georgi Gerganov] rolled in. He released llama.cpp on GitHub, which runs the inference of a LLaMa model with 4-bit quantization. His code was focused on running LLaMa-7B on your Macbook, but we’ve seen versions running on smartphones and Raspberry Pis. There’s even a version written in Rust! A rough rule of thumb is anything with more than 4 GB of RAM can run LLaMa. Model weights are available through Meta with some rather strict terms, but they’ve been leaked online and can be found even in a pull request on the GitHub repo itself. Continue reading “Why LLaMa Is A Big Deal” →

ChatGPT, Bing, And The Upcoming Security Apocalypse

March 4, 2023 by Elliot Williams 46 Comments

Most security professionals will tell you that it’s a lot easier to attack code systems than it is to defend them, and that this is especially true for large systems. The white hat’s job is to secure each and every point of contact, while the black hat’s goal is to find just one that’s insecure.

Whether black hat or white hat, it also helps a lot to know how the system works and exactly what it’s doing. When you’ve got the source code, either because it’s open-source, or because you’re working inside the company that makes the software, you’ve got a huge advantage both in finding bugs and in fixing them. In the case of closed-source software, the white hats arguably have the offsetting advantage that they at least can see the source code, and peek inside the black box, while the attackers cannot.

Still, if you look at the number of security issues raised weekly, it’s clear that even in the case of closed-source software, where the defenders should have the largest advantage, that offense is a lot easier than defense.

So now put yourself in the shoes of the poor folks who are going to try to secure large language models like ChatGPT, the new Bing, or Google’s soon-to-be-released Bard. They don’t understand their machines. Of course they know how the work inside, in the sense of cross multiplying tensors and updating weights based on training sets and so on. But because the billions of internal parameters interact in incomprehensible ways, almost all researchers refer to large language models’ inner workings as a black box.

And they haven’t even begun to consider security yet. They’re still worried about how to construct obscure background prompts that prevent their machines from spewing hate speech or pornographic novels. But as soon as the machines start doing something more interesting than just providing you plain text, the black hats will take notice, and someone will have to figure out defense.

Indeed, this week, we saw the first real shot across the bow: a hack to make Bing direct users to arbitrary (bad) webpages. The Bing hack requires the user to already be on a compromised website, so it’s maybe not very threatening, but it points out a possible real security difference between Bing and ChatGPT: Bing gives you links to follow, and that makes it a juicy target.

We’re right on the edge of a new security landscape, because even the white hats are facing a black box in the AI. So far, what ChatGPT and Codex and other large language models are doing is trivially secure – putting out plain text – but Bing is taking the first dangerous steps into doing something more useful, both for users and black hats. Given the ease with which people have undone OpenAI’s attempts to keep ChatGPT in its comfort zone, my guess is that the white hats will have their hands full, and the black-box nature of the model deprives them of their best hope. Buckle your seatbelts.