Update on Reflection-70B

https://glaive.ai/blog/post/reflection-postmortem

by mellosouls • 2 years ago

82 47 2 years ago

kristianp• 2 years ago

If this is for real, in some ways it shows how small OpenAIs moat is. Once someone knows something is possible and the rough idea, the community can replicate it in 4 weeks.

jsheard• 2 years ago

Isn't Reflection supposed to be based on CoT like o1? It was originally released a week before o1 was, so if it was the real deal all along then OpenAI were outright beaten to the punch rather than replicated after the fact.

ipsum25• 2 years ago

No. CoT has been around for several years (Jan 2022) https://arxiv.org/abs/2201.11903. And so has Reflection (March 2023) https://arxiv.org/abs/2303.11366. This approach that was taken by Reflection is nothing new.

nisten5• 2 years ago

yes and the massive increase to GPQA scores from o1 was attributed to this technique, so there is something there, despite the hard feelings of unproductive reddit users

thorum2• 2 years ago

CoT moderately improves model performance, but all non-o1 models suck at actually thinking step by step effectively. If the task is not straightforward, they make obvious mistakes or default to guessing.

OpenAI trained o1 to pick better steps in its chain of thought. (The moat is the dataset they used to do that.)

bastawhiz5• 2 years ago

Maybe, if it wasn't an outright fraud. Arguably they didn't beat anyone to anything.

coolspot• 2 years ago

On one hand, I want to believe Sahil, on the other hand most of his explanations don’t make much sense:

Can’t upload exact weights he had on his computer. The guy runs AI hosting/inference/training company - can’t upload weights he has!

Original benchmark harness wasn’t shared, but had a bug that conveniently boosted model results.

API somehow mysteriously censors model name and tokenizer is exact match to Claude.

all2• 2 years ago

I've seen stuff like this hacked together. If he isn't very organized or was hasty, there's a good bet he deleted the working weights or doesn't know which of 5 or 10 the weights it is.

Nothing would stop him from uploading all the weights, I suppose...

ipsum25• 2 years ago

No. He served the "weights" (actually Claude) for over 24 hours. It's practically impossible to have served the "correct weights" and just have lost them.

ameliaquining• 2 years ago

He seems to be claiming that anyone can now reproduce the weird Claude censorship locally with the uploaded weights. Has anyone checked whether that's true or not, or is he mischaracterizing the allegations?

xena2• 2 years ago

I'm going to be downloading the weights and doing local verification

ipsum25• 2 years ago

Sahil Chaudhary of GlaiveAI perpetuated fraud, where he replaced the model that he "trained" with other backend ML providers. He still has not given a reason why "Claude" the string would be missing, just magically happened, despite the base model, Llama3.1-70B having no issues producing the text "Claude" nor the dataset missing the string "Claude"!

Note that there was additional proof, besides missing the string "Claude", by matching the max number of tokens the model was able to produce. This is more technical, but chatGPT, Claude, Llama all have different tokenizers, so words are be broken up into different sections. The API consistently did NOT match the base model tokenizer (Llama), instead, producing the same number of tokens as Claude.

Companies and individuals should probably avoid GlaiveAI and Matt Shumer less they get scammed too.

nisten5• 2 years ago

they're 140Gb folders, for each checkpoint, yes file corruption happens

and as for the fraud part...it was an opensource model release that did not meet the claimed benchmarks when people tried to replicate it

alsodumb3• 2 years ago

Are you telling me someone trained a huge model, and served it for hours to tons of users, and had only one instance of the checkpoint? I call BS.

The model being open-source doesn't mean what they could have gotten away with, or tried to, isn't fraud.

bastawhiz5• 2 years ago

The fraud part was multiple independent sources producing fairly indisputable evidence that their "hosted" version of the model was just running GPT and Claude. That alone is enough to completely discredit absolutely everything about this work.

As for corruption, I don't believe the excuse "yes file corruption happens". They're model weights. If this was trained (in real life) it was done on some serious hardware with disks with error correction. They weren't storing the checkpoints on microSD cards. It's certainly possible that there was really unfortunate luck and there was corruption, but I don't find that excuse to be plausible. Especially when this is your business (and launch!)

ipsum25• 2 years ago

Definition of fraud, from Google:

* wrongful or criminal deception intended to result in financial or personal gain.

* a person or thing intended to deceive others, typically by unjustifiably claiming or being credited with accomplishments or qualities.

Since they were advertising GlaiveAI as this magical source of data where they trained a model that performed better than Claude and chatGPT, I think this firmly falls into that camp! Your definitions may be different than mine.

lolinder2• 2 years ago

Sorry, I'm having trouble finding more information about this—what is the significance of the model being unable to produce the string "Claude"? Was this some sort of half-hearted censorship to prevent it from disclosing its name? Where can I read more?

ipsum25• 2 years ago

https://x.com/shinboson/status/1832933747529834747 is the best summary.

if you don't have a twitter account: https://threadreaderapp.com/thread/1832933768031588622.html

0cf8612b2e1e• 2 years ago

Here was a HN post which links to a Reddit thread about this. Multiple users find evidence that the model was a lie. The most damning to me was that the tokenizer was not Llama’s, which cannot be explained away.

https://news.ycombinator.com/item?id=41484981

thorum2• 2 years ago

This along with a few tokenizer related tests people ran, made people suspect that we are just serving Claude with post-processing where we filter out words like Claude.

Didn't these "few tokenizer related tests" prove the API was using Claude's tokenizer instead of Llama's, based on how words were being divided into tokens?

That's a hard one to explain (it doesn't appear they're even trying to).

refulgentis4• 2 years ago

People keep asserting that but, really, it was just people pointing to setting max tokens to a certain value and getting a certain # of words out. They didn't actually have tokens. Perfectly possible to have collisions, I'd wager even likely in the scenarios they tested, simple question, < 10 tokens, in English.

blackeyeblitzar• 2 years ago

Past discussions about this model and its controversies:

Reflection 70B, the top open-source model https://news.ycombinator.com/item?id=41459781

Confirmed: Reflection 70B's official API is a wrapper for Sonnet 3.5 https://news.ycombinator.com/item?id=41484981

nisten5• 2 years ago

There is no official API, you're confirming a temporary one that was taken down weeks ago. That was done via OpenRouter, a routing api service/site, which routes to different model's under load.

Yes they could've switched it themselves too.

ilaksh• 2 years ago

The models are very powerful. This can help anyone, including scammers. The number of scams will be enormous.

I just hope people are also able to see how useful it is for non-scumbags and don't let the scammers ruin it for everyone else. I am not going to mention another similar category of technology in this regard, just to stay "politically correct" for this site.

talldayo• 2 years ago

I just hope people are also able to see how useful it is for non-scumbags and don't let the scammers ruin it for everyone else.

I hope so too - it's been four years since GPT-3 came out and I haven't found a single serious time-saving application for the technology.

If someone doesn't start making money with LLMs soon, then it will only be the scammers who benefit!

m3kw9• 2 years ago

Is either fraud or incompetency, I say is just incompetency like what he said. Got too excited on some false test maybe they tested with some validation data mixed in

bastawhiz5• 2 years ago

So when they put the hosted model online (which was actually just proxying Claude), they explicitly prompted Claude to censor its own name. That's not explainable with incompetence. It's very intentional deception.

jazzyjackson• 2 years ago

I just can't facepalm enough seeing so called AI companies relying on pythons .replace() when they need to hide what service they're building on

Havoc2• 2 years ago

An expensive lesson in how fragile reputations can be

alsodumb3• 2 years ago

I don't trust Sahil and Matt. They tried to commit fraud, hype things up, but it got way too much attention than what they expected, so they tried to get away with just serving Claude/ChatGPT in the background but got caught. They are nothing but grifters who got caught and now trying to fix that image.

bhouston2• 2 years ago

I am confused. He was hosting the 70B LLM everyone was demoing from his laptop? How can that serve the load? When I’ve run LLMs locally it is really taxing for just one concurrent session.

nisten5• 2 years ago

Has anyone here actually ran the code on their own hardware yet?

I did a standard non-middleware lm_eval_harness and got 0.3214 on gpqa_main_zeroshot WITH the systemprompt and 0.3616 without the systemprompt.

Have not ran it with the middleware yet that's supposed to do the substraction. Now, if that adds 20% to the score, that would be a huge deal, but it would also roughly match the jump from gpt4o to o1-preview that they got in gpqa_diamond.

Update on Reflection-70B

Related Stories