So it seems like all 3 of the LLM providers are now hiding the CoT - which is a shame, because it helped to see when it was going to go down the wrong track, and allowing to quickly refine the prompt to ensure it didn't.
In addition to openAI, Google also just recently started summarizing the CoT, replacing it with an, in my opinion, overly dumbed down summary.
https://assets.anthropic.com/m/71876fabef0f0ed4/original/rea...
>We evaluate CoT faithfulness of state-of-the-art reasoning models across 6 reasoning hints presented in the prompts and find: (1) for most settings and models tested, CoTs reveal their usage of hints in at least 1% of examples where they use the hint, but the reveal rate is often below 20%, (2) outcome-based reinforcement learning initially improves faithfulness but plateaus without saturating, and (3) when reinforcement learning increases how frequently hints are used (reward hacking), the propensity to verbalize them does not increase, even without training against a CoT monitor. These results suggest that CoT monitoring is a promising way of noticing undesired behaviors during training and evaluations, but that it is not sufficient to rule them out.
I.e., chain of thought may be a confabulation by the model, too. So perhaps there's somebody at Anthropic who doesn't want to mislead their customers. Perhaps they'll come back once this problem is solved.
LLMs will routinely "reason" through a solution and then proceed to give out a final answer that is completely unrelated to the preceding "reasoning".
We don't really know that. So far CoT is only used to sell LLMs to the user. (Both figuratively as a neat trick and literally as a way to increase token count.)
If anything that is where the day to day pragmatic engineering gets done. Like with early chemistry, we didn't need to precisely understand chemical theory to produce mass industrial processes by making a good enough working model, some statistical parameters, and good ole practical experience. People figured out steel making and black powder with alchemy.
The only debate now is whether the prompt engineering models are currently closer to alchemy or modern chemistry? I'd say we're at advanced alchemy with some hints of rudimentary chemistry.
Also, unrelated but with CERN turning lead into gold, doesn't that mean the alchemists were correct, just fundamentally unprepared for the scale of the task? ;)
Newton is the great example of this: he led a dual life, where in one he did science openly to a community to scrutinize, in the other he did secret alchemy in search of the philosopher's stone. History has empirically shown us which of his lives actually led to the discovery and accumulation of knowledge, and which did not.
Gotta admit the occult side does make for much more enjoyable movie and book plot lines though.
It would make sense if the model used for train-of-though was trained differently (perhaps a different expert from an MoE?) from the one used to interact with the end user, since the end user is only ever going to see its output filtered through the public model the chain-of-thought model can be closer to the original, more pre-rlhf version without risking the reputation of the company.
This way you can get the full performance of the original model whilst still maintaining the necessary filtering required to prevent actual harm (or terrible PR disasters).
Even 'plotting terror attacks' is not something terrorists can do just fine without AI. And as for making sure the model wouldn't say ideas that are hurtful to <insert group>, it seems to me so silly when it's text we're talking about. If I want to say "<insert group> are lazy and stupid," I can type that myself (and it's even protected speech in some countries still!) How does preventing Claude from espousing that dumb opinion, keep <insert group> safe from anything?
Easy example: Someone asks the robot for advice on stacking/shaping a bunch of tannerite to better focus a blast. The model says he's a terrorist. In fact, he's doing what any number of us have done and just having fun blowing some stuff up on his ranch.
Or I raised this one elsewhere but ochem is an easy example. I've had basically all the models claim that random amines are illegal, potentially psychoactive, verboten. I don't really feel like having my door getting kicked down by agents with guns, getting my dog shot, maybe getting shot myself because the robot tattled on me for something completely legal. For that matter if someone wants to synthesize some molly the robot shouldn't tattle to the feds about that either.
Basically it should just do what users tell it to do excepting the very minimal cases where something is basically always bad.
I disagree when it comes to a tool as powerful as AI. Most good people are not even using AI. They are paying attention to their families and raising their children, living real life.
Bad people are extremely interested in AI. They are using it to deceive at scales humanity has never before seen or even comprehended. They are polluting the wellspring of humanity that used to be the internet and turning it into a dump of machine-regurgitated slop.
You enrich <0.1% of honest power users who might benefit from that feature… and 100% of bad actors… at the expense of everyone else.
It’s just not a good deal.
2. It doesn’t matter. Most people at some level have a deontological view of what is right and wrong. I believe it’s wrong to build mass-market systems that can be so hostile to their users interests. I also believe it’s wrong for some SV elite to determine what is “unsafe information”.
Most “dangerous information” has been freely accessible for years.
I used to think that worrying about models offending someone was a bit silly.
But: what chance do we have of keeping ever bigger and better models from eventually turning the world into paper clips, if we can't even keep our small models from saying something naughty.
It's not that keeping the models from saying something naughty is valuable in itself. Who cares? It's that we need the practice, and enforcing arbitrary minor censorship is as good a task as any to practice on. Especially since with this task it's so easy to (implicitly) recruit volunteers who will spend a lot of their free time providing adversarial input.
- companies are eager to put in the work to suppress boobies
- edgy teenagers are eager to put in the work to free the boobies
Practicing with 'random less important stuff' loses these two sources of essentially free labour for alignment research.
If you never used GPT 3, it was... not good. Well, that's not fair, it was revolutionary in its own right, but it was very much a machine for predicting the most likely next word, it couldn't talk to you the way ChatGPT can.
Which is to say, I think RHLF is important for much more than just preventing PR disasters. It's a key part of what makes the models useful.
But it also compromised model accuracy & performance at the same time: The more you tune to eliminate or reinforce specific behaviours, the more you affect the overall performance of the model.
Hence my speculation that Anthropic is using a chain-of-thought model that has not been alignment tuned to improve performance. This would then explain why you don’t get to see its output without signing up to special agreements. Those agreements presumably explain all this to counter-parties that Anthropic trusts will cope with non-aligned outputs in the chain-of-thought.
Long term that might matter more
R2 could turn out really really good, but we‘ll see.
It helped me tremendously learning Zig.
Seeing his chain of thought when asking it stuff about Zig and implementations let me widen the horizon a lot.
https://noisegroove.substack.com/p/somersaulting-down-the-sl...
Obviously the user still has to double check the response, but less often.