Everyone Is Judging AI by These Tests. But Experts Say They’re Close to Meaningless.
Everyone Is Judging AI by These Tests. But Experts Say They’re Close to Meaningless.
Benchmarks used to rank AI models are several years old, often sourced from amateur websites, and, experts worry, lending automated systems a dubious sense of authority
Goodhart's law:
When a measure becomes a target, it ceases to be a good measure.
The Turing Test (as some people believe it to be): if you can have a conversation with a computer and not tell if it's a computer, then it must be intelligent.
AI companies: writes ML model that is specifically designed to convincingly play one side of a conversation, even though it has no ability to understand the things it talks about.
The most advanced models absolutely have modeling about what's being discussed and relationships between concepts.
Even toy models have been shown to build world models from very basic training data.
Honestly, read at least a little bit of the relevant research:
There's a reason why the open llm leaderboard was changed a while ago.
Basically, scores didn't improve much anymore and many tests were contained in the training data.See this blogpost for more info.
Much like IQ tests for humans are flawed too. Figuring out series of numbers or relations in a graphic representation, only tells how good you are at these specific tasks, and doesn't provide a reliable picture of "general" intelligence.
"close to meaningless" sums up my expert opinion on the whole current AI hype machine sales pitch.
Highly tuned models for incredibly specific, not-dangerous use cases is the next pragmatic step. There's a lot to excited about, in that very narrow band.
Anyone selling more than that is part of a con, or in very rare cases, doing genuine "fuck off and ask me again in a decade" kinds of research.
Looks quite satisfying to me, otherwise, we can still create new tests ... :
The tests cover an astounding range of knowledge, such as eighth-grade math, world history, and pop culture. Many are multiple choice, others take free-form answers. Some purport to measure knowledge of advanced fields like law, medicine and science. Others are more abstract, asking AI systems to choose the next logical step in a sequence of events, or to review “moral scenarios” and decide what actions would be considered acceptable behavior in society today.
This is the way:
The article makes the valid argument that LLMs simply predict next letters based on training and query.
But is that actually true of latest models from OpenAI, Claude etc?
And even if it is true, what solid proof do we have that humans aren’t doing the same? I’ve met endless people who could waffle for hours without seeming to do any reasoning.
Information theory, entropy in Markovian processes. Read up on these buzzwords to see why.
I think I know enough about these concepts to know that there isn’t any conclusive proof, observed in output or system state, to establish consensus that human speech output is generated differently to how LLMs generate output. If you have links to any papers that claim otherwise, I’ll be happy to read them.
what solid proof do we have that humans aren’t doing the same?
Humans are not computers. Brains are not LLMs...
Given a totally reasonable hypothesis (humans =/= computers) and a completely outlandish hypothesis (humans = computers), I would need much more 'proof' for the later.
Well, brains are a network of neurons (we can evidentially verify this) trained on … eyes, ears, sense of touch, taste, smell and balance (rewarded by endorphins released by the old brain on certain hardcoded stimuli). LLMs are a network of neurons trained on text and images (rewarded by producing text that mimics input text and some reasoning tests).
It’s not given that this results in the same way of dealing with language, given the wider set of input data for a human, but it’s not given that it doesn’t either.