BigMuffin69

has data access to much but not all of the dataset.

Huh! I wonder what part of the dset had the 25% of questions they got right in it 🙃

I can't believe they fucking got me with this one. I remember back in August(?) Epoch was getting quotes from top mathematicians like Tarrence Tao to review the benchmark and he was quoted saying like it would be a big deal for a model to do well on this benchmark, it will be several years before a model can solve all these questions organically etc so when O3 dropped and got a big jump from SotA, people (myself) were blown away. At the same time red flags were going up in my mind: Epoch was yapping about how this test was completely confidential and no one would get to see their very special test so the answers wouldn't get leaked. But then how in the hell did they evaluate this model on the test? There's no way O3 was run locally by Epoch at ~$1000 a question -> OAI had to be given the benchmark to run against in house -> maybe they had multiple attempts against it and were tuning the model/ recovering questions from api logs/paying mathematicians in house to produce answers to the problems so they could generate their own solution set??

No. The answer is much stupider. The entire company of Epoch ARE mathematicians working for OAI to make marketing grift to pump the latest toy. They got me lads, I drank the snake oil prepared specifically for people like me to drink :(

Reposting this for the new week thread since it truly is a record of how untrustworthy sammy and co are. Remember how OAI claimed that O3 had displayed superhuman levels on the mega hard Frontier Math exam written by Fields Medalist? Funny/totally not fishy story haha. Turns out OAI had exclusive access to that test for months and funded its creation and refused to let the creators of test publicly acknowledge this until after OAI did their big stupid magic trick.

From Subbarao Kambhampati via linkedIn:

"𝐎𝐧 𝐭𝐡𝐞 𝐬𝐞𝐞𝐝𝐲 𝐨𝐩𝐭𝐢𝐜𝐬 𝐨𝐟 “𝑩𝒖𝒊𝒍𝒅𝒊𝒏𝒈 𝒂𝒏 𝑨𝑮𝑰 𝑴𝒐𝒂𝒕 𝒃𝒚 𝑪𝒐𝒓𝒓𝒂𝒍𝒍𝒊𝒏𝒈 𝑩𝒆𝒏𝒄𝒉𝒎𝒂𝒓𝒌 𝑪𝒓𝒆𝒂𝒕𝒐𝒓𝒔” hashtag#SundayHarangue. One of the big reasons for the increased volume of “𝐀𝐆𝐈 𝐓𝐨𝐦𝐨𝐫𝐫𝐨𝐰” hype has been o3’s performance on the “frontier math” benchmark–something that other models basically had no handle on.

We are now being told (https://lnkd.in/gUaGKuAE) that this benchmark data may have been exclusively available (https://lnkd.in/g5E3tcse) to OpenAI since before o1–and that the benchmark creators were not allowed to disclose this *until after o3 *.

That o3 does well on frontier math held-out set is impressive, no doubt, but the mental picture of “𝒐1/𝒐3 𝒘𝒆𝒓𝒆 𝒋𝒖𝒔𝒕 𝒃𝒆𝒊𝒏𝒈 𝒕𝒓𝒂𝒊𝒏𝒆𝒅 𝒐𝒏 𝒔𝒊𝒎𝒑𝒍𝒆 𝒎𝒂𝒕𝒉, 𝒂𝒏𝒅 𝒕𝒉𝒆𝒚 𝒃𝒐𝒐𝒕𝒔𝒕𝒓𝒂𝒑𝒑𝒆𝒅 𝒕𝒉𝒆𝒎𝒔𝒆𝒍𝒗𝒆𝒔 𝒕𝒐 𝒇𝒓𝒐𝒏𝒕𝒊𝒆𝒓 𝒎𝒂𝒕𝒉”–that the AGI tomorrow crowd seem to have–that 𝘖𝘱𝘦𝘯𝘈𝘐 𝘸𝘩𝘪𝘭𝘦 𝘯𝘰𝘵 𝘦𝘹𝘱𝘭𝘪𝘤𝘪𝘵𝘭𝘺 𝘤𝘭𝘢𝘪𝘮𝘪𝘯𝘨, 𝘤𝘦𝘳𝘵𝘢𝘪𝘯𝘭𝘺 𝘥𝘪𝘥𝘯’𝘵 𝘥𝘪𝘳𝘦𝘤𝘵𝘭𝘺 𝘤𝘰𝘯𝘵𝘳𝘢𝘥𝘪𝘤𝘵–is shattered by this. (I have, in fact, been grumbling to my students since o3 announcement that I don’t completely believe that OpenAI didn’t have access to the Olympiad/Frontier Math data before hand… )

I do think o1/o3 are impressive technical achievements (see https://lnkd.in/gvVqmTG9 )

𝑫𝒐𝒊𝒏𝒈 𝒘𝒆𝒍𝒍 𝒐𝒏 𝒉𝒂𝒓𝒅 𝒃𝒆𝒏𝒄𝒉𝒎𝒂𝒓𝒌𝒔 𝒕𝒉𝒂𝒕 𝒚𝒐𝒖 𝒉𝒂𝒅 𝒑𝒓𝒊𝒐𝒓 𝒂𝒄𝒄𝒆𝒔𝒔 𝒕𝒐 𝒊𝒔 𝒔𝒕𝒊𝒍𝒍 𝒊𝒎𝒑𝒓𝒆𝒔𝒔𝒊𝒗𝒆–𝒃𝒖𝒕 𝒅𝒐𝒆𝒔𝒏’𝒕 𝒒𝒖𝒊𝒕𝒆 𝒔𝒄𝒓𝒆𝒂𝒎 “𝑨𝑮𝑰 𝑻𝒐𝒎𝒐𝒓𝒓𝒐𝒘.”

We all know that data contamination is an issue with LLMs and LRMs. We also know that reasoning claims need more careful vetting than “𝘸𝘦 𝘥𝘪𝘥𝘯’𝘵 𝘴𝘦𝘦 𝘵𝘩𝘢𝘵 𝘴𝘱𝘦𝘤𝘪𝘧𝘪𝘤 𝘱𝘳𝘰𝘣𝘭𝘦𝘮 𝘪𝘯𝘴𝘵𝘢𝘯𝘤𝘦 𝘥𝘶𝘳𝘪𝘯𝘨 𝘵𝘳𝘢𝘪𝘯𝘪𝘯𝘨” (see “In vs. Out of Distribution analyses are not that useful for understanding LLM reasoning capabilities” https://lnkd.in/gZ2wBM_F ).

At the very least, this episode further argues for increased vigilance/skepticism on the part of AI research community in how they parse the benchmark claims put out commercial entities."

Big stupid snake oil strikes again.

Remember how OAI claimed that O3 had displayed superhuman levels on the mega hard Frontier Math exam written by Fields Medalist? Funny/totally not fishy story haha. Turns out OAI had exclusive access to that test for months and funded its creation and refused to let the creators of test publicly acknowledge this until after OAI did their big stupid magic trick.

From Subbarao Kambhampati via linkedIn:

"𝐎𝐧 𝐭𝐡𝐞 𝐬𝐞𝐞𝐝𝐲 𝐨𝐩𝐭𝐢𝐜𝐬 𝐨𝐟 "𝑩𝒖𝒊𝒍𝒅𝒊𝒏𝒈 𝒂𝒏 𝑨𝑮𝑰 𝑴𝒐𝒂𝒕 𝒃𝒚 𝑪𝒐𝒓𝒓𝒂𝒍𝒍𝒊𝒏𝒈 𝑩𝒆𝒏𝒄𝒉𝒎𝒂𝒓𝒌 𝑪𝒓𝒆𝒂𝒕𝒐𝒓𝒔" hashtag#SundayHarangue. One of the big reasons for the increased volume of "𝐀𝐆𝐈 𝐓𝐨𝐦𝐨𝐫𝐫𝐨𝐰" hype has been o3's performance on the "frontier math" benchmark--something that other models basically had no handle on.

We are now being told (https://lnkd.in/gUaGKuAE) that this benchmark data may have been exclusively available (https://lnkd.in/g5E3tcse) to OpenAI since before o1--and that the benchmark creators were not allowed to disclose this *until after o3 *.

That o3 does well on frontier math held-out set is impressive, no doubt, but the mental picture of "𝒐1/𝒐3 𝒘𝒆𝒓𝒆 𝒋𝒖𝒔𝒕 𝒃𝒆𝒊𝒏𝒈 𝒕𝒓𝒂𝒊𝒏𝒆𝒅 𝒐𝒏 𝒔𝒊𝒎𝒑𝒍𝒆 𝒎𝒂𝒕𝒉, 𝒂𝒏𝒅 𝒕𝒉𝒆𝒚 𝒃𝒐𝒐𝒕𝒔𝒕𝒓𝒂𝒑𝒑𝒆𝒅 𝒕𝒉𝒆𝒎𝒔𝒆𝒍𝒗𝒆𝒔 𝒕𝒐 𝒇𝒓𝒐𝒏𝒕𝒊𝒆𝒓 𝒎𝒂𝒕𝒉"--that the AGI tomorrow crowd seem to have--that 𝘖𝘱𝘦𝘯𝘈𝘐 𝘸𝘩𝘪𝘭𝘦 𝘯𝘰𝘵 𝘦𝘹𝘱𝘭𝘪𝘤𝘪𝘵𝘭𝘺 𝘤𝘭𝘢𝘪𝘮𝘪𝘯𝘨, 𝘤𝘦𝘳𝘵𝘢𝘪𝘯𝘭𝘺 𝘥𝘪𝘥𝘯'𝘵 𝘥𝘪𝘳𝘦𝘤𝘵𝘭𝘺 𝘤𝘰𝘯𝘵𝘳𝘢𝘥𝘪𝘤𝘵--is shattered by this. (I have, in fact, been grumbling to my students since o3 announcement that I don't completely believe that OpenAI didn't have access to the Olympiad/Frontier Math data before hand.. )

I do think o1/o3 are impressive technical achievements (see https://lnkd.in/gvVqmTG9 )

𝑫𝒐𝒊𝒏𝒈 𝒘𝒆𝒍𝒍 𝒐𝒏 𝒉𝒂𝒓𝒅 𝒃𝒆𝒏𝒄𝒉𝒎𝒂𝒓𝒌𝒔 𝒕𝒉𝒂𝒕 𝒚𝒐𝒖 𝒉𝒂𝒅 𝒑𝒓𝒊𝒐𝒓 𝒂𝒄𝒄𝒆𝒔𝒔 𝒕𝒐 𝒊𝒔 𝒔𝒕𝒊𝒍𝒍 𝒊𝒎𝒑𝒓𝒆𝒔𝒔𝒊𝒗𝒆--𝒃𝒖𝒕 𝒅𝒐𝒆𝒔𝒏'𝒕 𝒒𝒖𝒊𝒕𝒆 𝒔𝒄𝒓𝒆𝒂𝒎 "𝑨𝑮𝑰 𝑻𝒐𝒎𝒐𝒓𝒓𝒐𝒘."

We all know that data contamination is an issue with LLMs and LRMs. We also know that reasoning claims need more careful vetting than "𝘸𝘦 𝘥𝘪𝘥𝘯'𝘵 𝘴𝘦𝘦 𝘵𝘩𝘢𝘵 𝘴𝘱𝘦𝘤𝘪𝘧𝘪𝘤 𝘱𝘳𝘰𝘣𝘭𝘦𝘮 𝘪𝘯𝘴𝘵𝘢𝘯𝘤𝘦 𝘥𝘶𝘳𝘪𝘯𝘨 𝘵𝘳𝘢𝘪𝘯𝘪𝘯𝘨" (see "In vs. Out of Distribution analyses are not that useful for understanding LLM reasoning capabilities" https://lnkd.in/gZ2wBM_F ).

At the very least, this episode further argues for increased vigilance/skepticism on the part of AI research community in how they parse the benchmark claims put out commercial entities."

Big stupid snake oil strikes again.

Lmaou. "We need to alignment pill the Russian youth." Fast forward to the year 20XX and the haunted alignment pilled adults are now 'aligning' their bots to the world's top nuclear armed despot.

tony_soprano_how_could_this_happen.jpg (for some reason awful systems won't let me upload pictures anymore (ノಠ益ಠ)ノ)

Holy Moses in heaven, iirc both Sam and Dario have said that their urge to build the torment nexus came from being inspired by online RAT forums. Maybe alignment 'pilling' youths is counterproductive to human flourishing? As the LWers say, "update your priors fuckheads"

ong Yann LeCun was sharing this post too and i was shook that he was seeing quality shit post like this before me. We are not ready for whats coming next . jpg

Fellas, I was promised the first catastrophic AI event in 2024 by the chief doomers. There's only a few hours left to go, I'm thinking skynet is hiding inside the times square orb. Stay vigilant!

The ARC scores don't matter too much to me at 3k a problem. Like the original goal of the prize had a compute limit. You can't break that rule and then claim victory ( I mean I guess you can, but like not everyone is gonna be as wowed as xitter randos, ensemble methods were already hitting 80% + acc to francois )

And unfortunately, with Frontier math, the lack of transparency w.r.t. which problems were solved and how they were solved makes it frustrating as hell to me, as someone who actually would like to see a super math robot. According to the senior math advisor to the people who created the data set, iirc 40% solved problems were in the easiest category / 50% in the second tier category and 10% were in the "hard" tier, but he said that he looked at the solutions and that they looked like mostly being solved 'heuristically' instead of plopping out any 'new' insights.

Again, none of this is good science, just pure shock and awe. I've heard rumors that OAI is hiring strong competition style mathematicians to supervise the reinforcement learning for these types of problems and if they are letting O3 take the test, then how the hell does that not leak the problem set? Like now the whole test is compromised now right? Since this behemoth uses enough electricity to power a city block, theres no way they would be able to run it locally. Now OAI can literally pay their peeps to solve the rest and surprise surprise O3++ will hit 80%

OTOH, with code forces scores and math scores this high, I can now put on my LW cap and say this model has 2 trillion IQ, so why hasn't it exterminated me and my family yet like big Yud promised? It's almost as if there is no little creature inside trying to take over the world or something.

Pat walking into the last board meeting

I remember when several months (a year ago?) when the news got out that gpt-3.5-turbo-papillion-grumpalumpgus could play chess around ~1600 elo. I was skeptical the apparent skill wasn't just a hacked-on patch to stop folks from clowning on their models on xitter. Like if an LLM had just read the instructions of chess and started playing like a competent player, that would be genuinely impressive. But if what happened is they generated 10^12 synthetic games of chess played by stonk fish and used that to train the model- that ain't an emergent ability, that's just brute forcing chess. The fact that larger, open-source models that perform better on other benchmarks, still flail at chess is just a glaring red flag that something funky was going on w/ gpt-3.5-turbo-instruct to drive home the "eMeRgEnCe" narrative. I'd bet decent odds if you played with modified rules, (knights move a one space longer L shape, you cannot move a pawn 2 moves after it last moved, etc), gpt-3.5 would fuckin suck.

Edit: the author asks "why skill go down tho" on later models. Like isn't it obvious? At that moment of time, chess skills weren't a priority so the trillions of synthetic games weren't included in the training? Like this isn't that big of a mystery...? It's not like other NN haven't been trained to play chess...

If they do press conferences this time around, ever question should just be "does Elon approve of decision ____ ?" Will drive Trump fkn insane.

I voted for Liz in 2020 :( instead they gave me diamond joe