AI

OpenAI’s new tool attempts to explain language models’ behaviors

Comment

OpenAI's logo
Image Credits: OpenAI

It’s often said that large language models (LLMs) along the lines of OpenAI’s ChatGPT are a black box, and certainly, there’s some truth to that. Even for data scientists, it’s difficult to know why, always, a model responds in the way it does, like inventing facts out of whole cloth.

In an effort to peel back the layers of LLMs, OpenAI is developing a tool to automatically identify which parts of an LLM are responsible for which of its behaviors. The engineers behind it stress that it’s in the early stages, but the code to run it is available in open source on GitHub as of this morning.

“We’re trying to [develop ways to] anticipate what the problems with an AI system will be,” William Saunders, the interpretability team manager at OpenAI, told TechCrunch in a phone interview. “We want to really be able to know that we can trust what the model is doing and the answer that it produces.”

To that end, OpenAI’s tool uses a language model (ironically) to figure out the functions of the components of other, architecturally simpler LLMs — specifically OpenAI’s own GPT-2.

OpenAI explainability
OpenAI’s tool attempts to simulate the behaviors of neurons in an LLM. Image Credits: OpenAI

How? First, a quick explainer on LLMs for background. Like the brain, they’re made up of “neurons,” which observe some specific pattern in text to influence what the overall model “says” next. For example, given a prompt about superheros (e.g. “Which superheros have the most useful superpowers?”), a “Marvel superhero neuron” might boost the probability the model names specific superheroes from Marvel movies.

OpenAI’s tool exploits this setup to break models down into their individual pieces. First, the tool runs text sequences through the model being evaluated and waits for cases where a particular neuron “activates” frequently. Next, it “shows” GPT-4, OpenAI’s latest text-generating AI model, these highly active neurons and has GPT-4 generate an explanation. To determine how accurate the explanation is, the tool provides GPT-4 with text sequences and has it predict, or simulate, how the neuron would behave. In then compares the behavior of the simulated neuron with the behavior of the actual neuron.

“Using this methodology, we can basically, for every single neuron, come up with some kind of preliminary natural language explanation for what it’s doing and also have a score for how how well that explanation matches the actual behavior,” Jeff Wu, who leads the scalable alignment team at OpenAI, said. “We’re using GPT-4 as part of the process to produce explanations of what a neuron is looking for and then score how well those explanations match the reality of what it’s doing.”

The researchers were able to generate explanations for all 307,200 neurons in GPT-2, which they compiled in a dataset that’s been released alongside the tool code.

Tools like this could one day be used to improve an LLM’s performance, the researchers say — for example to cut down on bias or toxicity. But they acknowledge that it has a long way to go before it’s genuinely useful. The tool was confident in its explanations for about 1,000 of those neurons, a small fraction of the total.

A cynical person might argue, too, that the tool is essentially an advertisement for GPT-4, given that it requires GPT-4 to work. Other LLM interpretability tools are less dependent on commercial APIs, like DeepMind’s Tracr, a compiler that translates programs into neural network models.

Wu said that isn’t the case — the fact the tool uses GPT-4 is merely “incidental” — and, on the contrary, shows GPT-4’s weaknesses in this area. He also said it wasn’t created with commercial applications in mind and, in theory, could be adapted to use LLMs besides GPT-4.

OpenAI explainability
The tool identifies neurons activating across layers in the LLM. Image Credits: OpenAI

“Most of the explanations score quite poorly or don’t explain that much of the behavior of the actual neuron,” Wu said. “A lot of the neurons, for example, are active in a way where it’s very hard to tell what’s going on — like they activate on five or six different things, but there’s no discernible pattern. Sometimes there is a discernible pattern, but GPT-4 is unable to find it.”

That’s to say nothing of more complex, newer and larger models, or models that can browse the web for information. But on that second point, Wu believes that web browsing wouldn’t change the tool’s underlying mechanisms much. It could simply be tweaked, he says, to figure out why neurons decide to make certain search engine queries or access particular websites.

“We hope that this will open up a promising avenue to address interpretability in an automated way that others can build on and contribute to,” Wu said. “The hope is that we really actually have good explanations of not just what neurons are responding to but overall, the behavior of these models — what kinds of circuits they’re computing and how certain neurons affect other neurons.”

More TechCrunch

The European Union has warned Microsoft it could be fined up to 1% of its global annual turnover under the bloc’s online governance regime, the Digital Services Act (DSA), after…

Microsoft warned it could be fined billions by EU over missing GenAI risk info

The prospects for troubled banking-as-a-service startup Synapse have gone from bad to worse this week after a United States Trustee filed an emergency motion on Wednesday.  The trustee is asking…

A US Trustee wants troubled fintech Synapse to be liquidated via Chapter 7 bankruptcy, cites ‘gross mismanagement’

U.K.-based Seraphim Space is spinning up its 13th accelerator program, with nine participating companies working on a range of tech from propulsion to in-space manufacturing and space situational awareness. The…

Seraphim’s latest space accelerator welcomes nine companies

OpenAI has reached a deal with Reddit to use the social news site’s data for training AI models. In a blog post on OpenAI’s press relations site, the company said…

OpenAI inks deal to train AI on Reddit data

X users will now be able to discover posts from new Communities that are trending directly from an Explore tab within the section.

X pushes more users to Communities

For Mark Zuckerberg’s 40th birthday, his wife got him a photoshoot. Zuckerberg gives the camera a sly smile as he sits amid a carefully crafted re-creation of his childhood bedroom.…

Mark Zuckerberg’s makeover: Midlife crisis or carefully crafted rebrand?

Strava announced a slew of features, including AI to weed out leaderboard cheats, a new ‘family’ subscription plan, dark mode and more.

Strava taps AI to weed out leaderboard cheats, unveils ‘family’ plan, dark mode and more

We all fall down sometimes. Astronauts are no exception. You need to be in peak physical condition for space travel, but bulky space suits and lower gravity levels can be…

Astronauts fall over. Robotic limbs can help them back up.

Microsoft will launch its custom Cobalt 100 chips to customers as a public preview at its Build conference next week, TechCrunch has learned. In an analyst briefing ahead of Build,…

Microsoft’s custom Cobalt chips will come to Azure next week

What a wild week for transportation news! It was a smorgasbord of news that seemed to touch every sector and theme in transportation.

Tesla keeps cutting jobs and the feds probe Waymo

Sony Music Group has sent letters to more than 700 tech companies and music streaming services to warn them not to use its music to train AI without explicit permission.…

Sony Music warns tech companies over ‘unauthorized’ use of its content to train AI

Winston Chi, Butter’s founder and CEO, told TechCrunch that “most parties, including our investors and us, are making money” from the exit.

GrubMarket buys Butter to give its food distribution tech an AI boost

The investor lawsuit is related to Bolt securing a $30 million personal loan to Ryan Breslow, which was later defaulted on.

Bolt founder Ryan Breslow wants to settle an investor lawsuit by returning $37 million worth of shares

Meta, the parent company of Facebook, launched an enterprise version of the prominent social network in 2015. It always seemed like a stretch for a company built on a consumer…

With the end of Workplace, it’s fair to wonder if Meta was ever serious about the enterprise

X, formerly Twitter, turned TweetDeck into X Pro and pushed it behind a paywall. But there is a new column-based social media tool in town, and it’s from Instagram Threads.…

Meta Threads is testing pinned columns on the web, similar to the old TweetDeck

As part of 2024’s Accessibility Awareness Day, Google is showing off some updates to Android that should be useful to folks with mobility or vision impairments. Project Gameface allows gamers…

Google expands hands-free and eyes-free interfaces on Android

A hacker listed the data allegedly breached from Samco on a known cybercrime forum.

Hacker claims theft of India’s Samco account data

A top European privacy watchdog is investigating following the recent breaches of Dell customers’ personal information, TechCrunch has learned.  Ireland’s Data Protection Commission (DPC) deputy commissioner Graham Doyle confirmed to…

Ireland privacy watchdog confirms Dell data breach investigation

Ampere and Qualcomm aren’t the most obvious of partners. Both, after all, offer Arm-based chips for running data center servers (though Qualcomm’s largest market remains mobile). But as the two…

Ampere teams up with Qualcomm to launch an Arm-based AI server

At Google’s I/O developer conference, the company made its case to developers — and to some extent, consumers — why its bets on AI are ahead of rivals. At the…

Google I/O was an AI evolution, not a revolution

TechCrunch Disrupt has always been the ultimate convergence point for all things startup and tech. In the bustling world of innovation, it serves as the “big top” tent, where entrepreneurs,…

Meet the Magnificent Six: A tour of the stages at Disrupt 2024

There’s apparently a lot of demand for an on-demand handyperson. Khosla Ventures and Pear VC have just tripled down on their investment in Honey Homes, which offers up a dedicated…

Khosla Ventures, Pear VC triple down on Honey Homes, a smart way to hire a handyman

TikTok is testing the ability for users to upload 60-minute videos, the company confirmed to TechCrunch on Thursday. The feature is available to a limited group of users in select…

TikTok tests 60-minute video uploads as it continues to take on YouTube

Flock Safety is a multibillion-dollar startup that’s got eyes everywhere. As of Wednesday, with the company’s new Solar Condor cameras, those eyes are solar-powered and use wireless 5G networks to…

Flock Safety’s solar-powered cameras could make surveillance more widespread

Since he was very young, Bar Mor knew that he would inevitably do something with real estate. His family was involved in all types of real estate projects, from ground-up…

Agora raises $34M Series B to keep building the Carta for real estate

Poshmark, the social commerce site that lets people buy and sell new and used items to each other, launched a paid marketing tool on Thursday, giving sellers the ability to…

Poshmark’s ‘Promoted Closet’ tool lets sellers boost all their listings at once

Google is launching a Gemini add-on for educational institutes through Google Workspace.

Google adds Gemini to its Education suite

More money for the generative AI boom: Y Combinator-backed developer infrastructure startup Recall.ai announced Thursday it has raised a $10 million Series A funding round, bringing its total raised to over…

YC-backed Recall.ai gets $10M Series A to help companies use virtual meeting data

Engineers Adam Keating and Jeremy Andrews were tired of using spreadsheets and screenshots to collab with teammates — so they launched a startup, CoLab, to build a better way. The…

CoLab’s collaborative tools for engineers line up $21M in new funding

Reddit announced on Wednesday that it is reintroducing its awards system after shutting down the program last year. The company said that most of the mechanisms related to awards will…

Reddit reintroduces its awards system