Anthropic made lots of intriguing discoveries using this approach, not least of which is why LLMs are so terrible at basic mathematics. "Ask Claude to add 36 and 59 and the model will go through a series of odd steps, including first adding a selection of approximate values (add 40ish and 60ish, add 57ish and 36ish). Towards the end of its process, it comes up with the value 92ish. Meanwhile, another sequence of steps focuses on the last digits, 6 and 9, and determines that the answer must end in a 5. Putting that together with 92ish gives the correct answer of 95," the MIT article explains.

But here's the really funky bit. If you ask Claude how it got the correct answer of 95, it will apparently tell you, "I added the ones (6+9=15), carried the 1, then added the 10s (3+5+1=9), resulting in 95." But that actually only reflects common answers in its training data as to how the sum might be completed, as opposed to what it actually did.

Another very surprising outcome of the research is the discovery that these LLMs do not, as is widely assumed, operate by merely predicting the next word. By tracing how Claude generated rhyming couplets, Anthropic found that it chose the rhyming word at the end of verses first, then filled in the rest of the line.

To understand what's actually happening, Anthropic's researchers developed a new technique, called circuit tracing, to track the decision-making processes inside a large language model step-by-step. They then applied it to their own Claude 3.5 Haiku LLM.

Anthropic says its approach was inspired by the brain scanning techniques used in neuroscience and can identify components of the model that are active at different times. In other words, it's a little like a brain scanner spotting which parts of the brain are firing during a cognitive process.

This is why LLMs are so patchy at math. (Image credit: Anthropic)

Anthropic made lots of intriguing discoveries using this approach, not least of which is why LLMs are so terrible at basic mathematics. "Ask Claude to add 36 and 59 and the model will go through a series of odd steps, including first adding a selection of approximate values (add 40ish and 60ish, add 57ish and 36ish). Towards the end of its process, it comes up with the value 92ish. Meanwhile, another sequence of steps focuses on the last digits, 6 and 9, and determines that the answer must end in a 5. Putting that together with 92ish gives the correct answer of 95," the MIT article explains.

But here's the really funky bit. If you ask Claude how it got the correct answer of 95, it will apparently tell you, "I added the ones (6+9=15), carried the 1, then added the 10s (3+5+1=9), resulting in 95." But that actually only reflects common answers in its training data as to how the sum might be completed, as opposed to what it actually did.

In other words, not only does the model use a very, very odd method to do the maths, you can't trust its explanations as to what it has just done. That's significant and shows that model outputs can not be relied upon when designing guardrails for AI. Their internal workings need to be understood, too.

Another very surprising outcome of the research is the discovery that these LLMs do not, as is widely assumed, operate by merely predicting the next word. By tracing how Claude generated rhyming couplets, Anthropic found that it chose the rhyming word at the end of verses first, then filled in the rest of the line.

"The planning thing in poems blew me away," says Batson. "Instead of at the very last minute trying to make the rhyme make sense, it knows where it’s going."

Anthropic discovered that their Claude LLM didn't just predict the next word. (Image credit: Anthropic)

Anthropic also found, among other things, that Claude "sometimes thinks in a conceptual space that is shared between languages, suggesting it has a kind of universal 'language of thought'."

Anywho, there's apparently a long way to go with this research. According to Anthropic, "it currently takes a few hours of human effort to understand the circuits we see, even on prompts with only tens of words." And the research doesn't explain how the structures inside LLMs are formed in the first place.

But it has shone a light on at least some parts of how these oddly mysterious AI beings—which we have created but don't understand—actually work. And that has to be a good thing.

'is weirder than you thought '

I am as likely to click a link with that line as much as if it had

'this one weird trick' or 'side hussle'.

I would really like it if headlines treated us like adults and got rid of click baity lines.

"Ask Claude to add 36 and 59 and the model will go through a series of odd steps, including first adding a selection of approximate values (add 40ish and 60ish, add 57ish and 36ish). Towards the end of its process, it comes up with the value 92ish. Meanwhile, another sequence of steps focuses on the last digits, 6 and 9, and determines that the answer must end in a 5. Putting that together with 92ish gives the correct answer of 95," the MIT article explains."

That is precisrly how I do math. Feel a little targeted that they called this odd.

The research paper looks well written but I couldn’t find any information on if this paper is going to be published in a reputable journal and peer reviewed. I have little faith in private businesses who profit from AI providing an unbiased view of how AI works. I think the first question I’d like answered is did Anthropic’s marketing department review the paper and did they offer any corrections or feedback? We’ve all heard the stories about the tobacco industry paying for papers to be written about the benefits of smoking and refuting health concerns.

this is one of the most interesting things about Llms that i have ever read

But here’s the really funky bit. If you ask Claude how it got the correct answer of 95, it will apparently tell you, “I added the ones (6+9=15), carried the 1, then added the 10s (3+5+1=9), resulting in 95.” But that actually only reflects common answers in its training data as to how the sum might be completed, as opposed to what it actually did.

This is not surprising. LLMs are not designed to have any introspection capabilities.

Introspection could probably be tacked onto existing architectures in a few different ways, but as far as I know nobody's done it yet. It will be interesting to see how that might change LLM behavior.

you can't trust its explanations as to what it has just done.

I might have had a lucky guess, but this was basically my assumption. You can't ask LLMs how they work and get an answer coming from an internal understanding of themselves, because they have no 'internal' experience.

Unless you make a scanner like the one in the study, non-verbal processing is as much of a black box to their 'output voice' as it is to us.

The other day I asked an llm to create a partial number chart to help my son learn what numbers are next to each other. If I instructed it to do this using very detailed instructions it failed miserably every time. And sometimes when I even told it to correct specific things about its answer it still basically ignored me. The only way I could get it to do what I wanted consistently was to break the instructions down into small steps and tell it to show me its pr.ogress.

I'd be very interested to learn it's "thought process" in each of those scenarios.

Another very surprising outcome of the research is the discovery that these LLMs do not, as is widely assumed, operate by merely predicting the next word. By tracing how Claude generated rhyming couplets, Anthropic found that it chose the rhyming word at the end of verses first, then filled in the rest of the line.

If the llm already knows the full sentence it's going to output from the first word it "guesses" I wonder if you could short circuit it and say just give the full sentence instead of doing a cycle for each word of the sentence, could maybe cut down on llm energy costs.

Wow, interesting. :)

Not unexpectedly, the LLM failed to explain its own thought process correctly.