- Tweets 3,558
- Following 97
- Followers 470
- Likes 5,721
Honestly I think I can do a decent postmortem of everything which went wrong with 5:
1. Two and a half years ago, long before they had any idea what it would look like, they were hyping 5.
2. They trained Orion as 5, expecting big benefits from scaling pretraining, but (1/x)
29
83
2,225
1,143,826
unexpectedly it turned out that the signs of diminishing returns were correct.
3. Lucky for them they got bailed out by the relatively unexpected success of RLVR and the O-series.
4. Just while they're dealing with Orion sucking compared to what they expected...(2/x)
2
6
399
50,423
scaling from o1 to o3 surprisingly works very well. They're saved! They're also led to believe that continued scaling of RLVR will keep delivering amazing results (fatal mistake).
5. Meanwhile Sama gets tired of being lightly made fun of for the proliferation of models (3/x)
1
4
328
47,109
The model router becomes a pet idea among executives and they decide to scrap the o3 release in favor of "GPT-5."
6. They've having more struggles with o3 in general then they're willing to admit and the router idea doesn't really work well (but it's a pet idea, so...) (4/x)
2
3
269
47,733
And then we end up at today. A lot of furious backtracking is happening, and their hyped release has gone poorly in every possible way (9/9).
1
7
265
33,975
All the while OAI employees are building a model for their own needs. How most people use and see their models really doesn't matter to them, and they apparently don't use the trial router internally. Meanwhile hype is turned up to 100% because that's the culture (7/x)
4
5
260
39,721
Things start looking grim in the summer. Nothing is as impressive as it should be. No big leap for GPT-5 is apparent. There is talk of delaying again. However Sama tries out the latest checkpoint and really loves it. With the cult leader/CEO directing, release is ordered (6/x).
2
5
256
39,879
They decide to release o3 by its lonesome and push "GPT-5" back a few months. They figure that with scaling RL they can deliver something really impressive by then. There are probably already signs that scaling RLVR is disappointing but they're ignored. (5/x)
1
5
240
42,291
Low energy and poorly planned release stream because despite the hype, no one other than Sama is actually passionate about this model in particular. Also lots of unilateral decision making for their customers and some poor UI and infrastructure planning, because of course (8/x)
1
5
229
36,063
Just to be clear, I don't think GPT-5 is some sort of awful model Llama 4 style, I just explaining why things have gone so poorly despite it being at the very least solid.
1
4
209
29,753
If they had delivered something GPT-OSS style (benchmaxed model of limited broad utility), they'd be in a far worse circumstance. That might be company-killing.
4
1
169
28,123
Generally, the students most privileged by the current American college admission system are academically mediocre students from affluent backgrounds. Institutions find admitting them lucrative since they receive little financial aid and pay close to the full cost of tuition.
3
7
113
10,145
In my opinion, few TV series have pulled off that kind of coherent arc since Babylon 5, too. They all claim to present long-term arcs but either the writers are making up everything as they go along (BSG 2003) or the show gets canceled before they can finish (Dark Matter).
12
1
97
87,323
I like DS9 but it's pretty evident that they were making stuff up as they went along and didn't have a plan until near the end. For Korra they would plan each season out ahead of time but they didn't really have a master plan for the whole show. Only ATLA seems comparable to B5.
2
50
4,162
"Israeli's strategy of killing and starving random civilians has been surprisingly ineffective at defeating Hamas, but the U.S. should keep unconditionally supporting it since..."
1
40
1,712
I noticed something similar with 4.1-mini too. Never seen a LLM output that answer before.
2
46
47,859
If the next Gemini is incredible, GPT-5 will look bad by comparison. If it's OK or bad, no one will care much. As for Claude vs. Gemini vs. GPT-5, probably depends on what you're using it for. Anyways, OpenAI depends on "hype" more than any other lab.
2
45
12,773
I don't know any countries that have regretted building nuclear weapons, but I know of many who have regretted not building them.
3
42
2,717
I think Nazi Germany was very good at fighting though, but very bad at grand strategy. They didn't conquer almost all of Europe through being bad at war. However, fascist ideology resulted in them biting off far more than they could chew and they lost.
16
1
40
18,471
It's 2035, and OpenAI is releasing GPT-4.9o-Turbo-Max-Plus. Devoted OA fanboys still claim that GPT-5 is "just around the corner."
2
23
3,682
The Expanse is like Game of Thrones, with the adaption element doing most of the writer's planning for them.
1
21
1,808
I would say that "AI" is an ideological project defined by its objective of "replacing" humans. That's what stands out to me most about the field. It even helps explain why the field is so obsessed with rejecting its past, since previous methods failed at "human-replacement."
6
4
23
749
Dortmund generated ~1.9 xG in 54 minutes with Götze playing and ~0.36 xG in 38 minutes without him playing, yet most Dortmund fans only say that he was "solid." #perception
1
8
19
I feel a little special for having known that Yudkowsky and LessWrong were complete garbage for around eight years. Harry Potter and the Methods of Rationality makes a wonderful case for not allowing its author to ever be in charge of anything at all.
1
24
2,808
A Weimar Republic trying to fight the Soviet Union, UK, and U.S. at the same time also would have lost, but a Weimar Republic would never have tried it in the first place.
1
22
1,507
What's your opinion on how reinforcement trained LLMs like R1 fit into this paradigm? They can definitely do at least some things beyond what ordinary LLMs can.
1
7
3,476
I've heard that units are reporting everyone they kill as "terrorists" to get their kill counts higher.
1
1
19
334
I mean that's also possible but we shouldn't underrate how dumb they are. They're perfectly possible of destroying social security because they thought deduplicating the database would stop fraud.
1
18
446
I think we're looking at the end of the open web in a few years at this rate--which is ironic, since Google's whole business model derives from its existence.
2
15
816
The fans of every team he's played for outside of Chelsea seem to agree he's a pretty good player, but that has zero chance of penetrating the Premier League discourse.
13
1,786
The thing is that I think that the vast majority of people are not OK with having their data being used to train AI systems. If we had had much more public awareness of what was of going on ~8 years ago, we might not be in the situation we are now.
1
9
829
@GaryMarcus You might find this interesting.
14
26,884
I can tell you too that OAI are not the only ones who built or played with a model router. xAI did it. The LMArena tested one. Switchpoint Router is a thing. I think Google tested one(?). OAI are just unique in making stripping control a "selling feature."
The model router becomes a pet idea among executives and they decide to scrap the o3 release in favor of "GPT-5."
6. They've having more struggles with o3 in general then they're willing to admit and the router idea doesn't really work well (but it's a pet idea, so...) (4/x)
1
14
2,318
It's been there since May 2023, at least: web.archive.org/web/20230517…
1
14
1,568
What the markets should be pricing in is that we're being ruled by a mad king who incompetently makes massive decisions based on a whim.
1
1
12
553
And apparently the Semianalysis report is badly, badly inaccurate anyways:
This key chart from @SemiAnalysis_ appears to have been the key source for claims of "50,000 Hoppers" and more detailed disclosure on their CapEx buildup analysis ("$1.3B").
But the table has errors/inconsistencies. More significantly, key assumptions don't pass sanity checks.
ALT https://semianalysis.com/2025/01/31/deepseek-debates/
2
13
1,332
I was literally trying to calculate the amount of "high quality" tokens in existence two years ago and wondering the same thing. All I can assume is that after GPT2->GPT3->GPT4 scaleups all worked, there wasn't a lot of room for doubt within the organization.
1
14
4,167
I looked up the "OpenAI researcher" that was cited here. He's not an AI researcher, he's a "philosopher" deeply in the LessWrong and EA cults who takes the "singularity" as a given. The fact that he's an employee at OpenAI is deeply indicative of the institutional culture there.
12
212
I'm not Jewish but I feel like I'm facing a similar crises about America right now, and if anything I'm becoming a stronger American patriot for it. There's something in America worth fighting for and I'll be damned if I let fascists and traitors define the country.
1
10
280
But why kill Navalny? I understand your point, but Navalny is by your criteria no threat at all. Hardly seems worth going through the effort to kill.
8
9
131,315
The entire "early model access" thing has always felt like a scam to me to be honest. You know they're generally only going to give the access to people who will hype, and those who end up being critical aren't going to get access in the future.
10
509
Doesn't this depend on the issue? I am all in favor of being understanding and compassionate, and I know that all people believe themselves to be reasonable, but I think that, for instance, that if you support mass slavery you cannot be a good person by my standards.
1
10
641
Yeah, but I'd never seen 0.31 as the answer before. Perhaps a side effect of mitigation efforts?
10
826
Supposedly, nearly half of the bombs they've dropped are unguided: cnn.com/2023/12/13/politics/… . I found that shocking when I read it because the primary reason for the adoption of "precision weapons" in the first place was that they were much more effective militarily.
8
606
They also don't really explain what they did differently? Entire thing smells. Anyways, now I'm tempted to take one of the world class open weight models, finetune for a five minutes on something that won't lobotomize it, and then release it as a "frontier model."
2
10
430
Somehow I have doubts about the loyalty and reliability of the algorithmically chosen security details.
1
9
251
The aviation industry has pushed this sort of automation as far as they could so far, occasionally with some very bad results.
2
46
If Britain had to pay reparations to all the countries and peoples they screwed over with their colonialism, the UK would be bankrupt. Well, more bankrupt. Israel/Palestine wasn't even their largest scale decolonializing screwup in 1947.
1
9
343
I noticed that Bayern lost the ability to play through Dortmund's press after Gotze came in; many more long balls to no one.
2
5
10
Either people will stop publishing information on the web(aside maybe from ultracheap AI contents farms), or they will publish it but take steps to prevent it from being cached by crawlers, so AI's can't plagiarize it.
6
140
In competive games (non-friendlies) where both Muller and Messi started(7 instances), Muller's team has scored 22, while Messi's team has scored 7. Muller has won 6 out of 7 of these games.
1
1
8
11,979
"The Holocaust wasn't especially unique"...dear god that's just essentially a basic description of reality. The Holocaust was barely even the biggest genocide perpetuated by Nazi Germany in particular. "Only" about a third of Nazi genocide victims were Jews.
1
9
225
same five problems solved, but google's proofs are actually readable and actually officially verified by the organizers. Same sixth problem failed. Unclear how different the approaches were.
9
747
Israel's actions indicate how terrified they are that HTS might actually set up a competent, effective government with popular legitimacy.
1
1
6
59
A couple months ago I saw someone tweet "how could LLMs answer questions about the world without a world model?" My instant thought was "doesn't Siri answer questions about the world?"
9
3,347
covid19.who.int/region/wpro/…
China has in fact reported ~1000 deaths(and ~50000 cases) since April 2020, so nothing you're saying has any basis in fact.
7
The rate limit for O1 in ChatGPT Plus is about 1/150th of the rate limit for 4o, when by API token pricing it should only be about 1/4th. Something seems to be making a lot of tokens disappear into thin air...
2
8
3,603
Historically the primary utility of computers has come from finding ways to use them to compliment humans and do things people are bad at or can't do at all(i.e. "The Mother of All Demos"). "AI" as a field is defined by its refusal to learn that "bitter lesson."
1
2
8
204
I think we have to assume that targeting civilians was at least to some degree preplanned, but the scale at which it ended up happening was not entirely planned. And I suspect Hamas didn't really have much plans for civilian hostages.
1
7
1,373
Lex Friedman literally just had an interview where Sam Altman implied that GPT-5 might not be released this year and they hadn't decided what to call their next LLM release in any case: lexfridman.com/sam-altman-2-…
1
1
6
18,319
Self-driving cars were "basically" solved in 1995: web.archive.org/web/20180612…
7
231
Someone trained an n-gram model on five trillion tokens: arxiv.org/abs/2401.17377
1
6
176
I live in a swing state, and that's like 90% of what the Harris ad spending is on. It's all majoritarian economic messaging. Honestly I would like to see a couple abortion or "Trump tried a coup" ads to work a different angle once in a while.
1
7
363
Wait, I didn't realize that AIME is literally just 15 problems every year. Why are people using a 15 problem set as a major benchmark?
2
7
1,657
v3 has like 95% of the performance for about 1% of the training cost and 1% of the inference cost. Sad!
1
8
1,122
If you're interested, you can watch the Chinese adaptation on several streaming services(Prime, Peacock, Viki, Plex) in the U.S. with subtitles: justwatch.com/us/tv-show/thr…. I have no idea how good it is but might be worth a shot.
3
1
6
603
This is justified as a means to balance the budget. However, even institutions with less budgetary concerns(i.e. Ivy League) find ways to reward this set of students, like legacy admissions.
1
4
720
I don't understand the obsession with using chatbots to replace search. Despite Google's obsession with continually making using search a worse experience, the current search web ecosystem generally works well and is remarkably versatile.
1
1
6
425
And of course "AI Safety" is largely about the desire to replace and disempower others while avoiding replacement yourself.
1
7
109
It's also always the elite in any given society which have the most ability and capacity to present their views and ideas to the "West." This greatly distorts "Western" perceptions and policies.
1
6
858
If you read the entire thread, I think there were a few deep sins involved in, notwithstanding technical problems:
1. Hype and committing to a model naming scheme based on it.
2. Sama making decisions based on not liking being made fun of on twitter (1/x)
Honestly I think I can do a decent postmortem of everything which went wrong with 5:
1. Two and a half years ago, long before they had any idea what it would look like, they were hyping 5.
2. They trained Orion as 5, expecting big benefits from scaling pretraining, but (1/x)
1
7
530
I don't think Golan Heights have ever been an important issues to the Palestinian leadership. However, Israel did ethnically cleanse the population there (except for Druze, who were perceived as "useful") and a peace treaty with Syria would require a settlement about Golan.
1
7
492
Not one of benchmarks, knowledge cutoff date, or context window were in the model card. They couldn't be bothered to include the most basic, utilitarian info on the model itself.
1
7
350
So the spammy websites over the last few years which felt like they were machine written might have actually been written by machines?
1
4
492
This sort of whining just blackpills me more about what the quality of Grok 3 is going to be.
1
6
289
"This can't be the right way to program a computer." --@colin_fraser
6
143
The training costs of "state of the art" deep learning models have increased something like six orders of magnitude over the last decade. Odds are that if Sama got 7 trillion, he would in a few years discover that he really needed 700 trillion, and if he got 700 trillion...
5
388
That's been my suspicion for a while now. Chinese industrial capacity is much larger and Chinese government and society seems significantly more dynamic and adaptable.
4
5
3,563
Babylon 5 had all of its mysteries planned out in advance, but it wasn't really a mystery box show. A lot of mysteries, but the answers to them were gradually revealed.
4
337
I'm not a huge fan of Hinton, but this burn is so sick he deserves another Noble for it.
I think Elon Musk should be expelled from the British Royal Society. Not because he peddles conspiracy theories and makes Nazi salutes, but because of the huge damage he is doing to scientific institutions in the US. Now let's see if he really believes in free speech.
1
4
224
At this point, I don't expect anything from model "technical papers" other than bragging about improvements on benchmarks and a handful of cherrypicked examples to serve as marketing. Of course, Google bragging about how they beat GPT-4 by 1% on benchmarks is particularly sad.
1
1
5
282
Your point about synthetic data is great, but always bear in mind that there are excellent odds that in 20 years time LLMs will not be considered "AI" at all, because the field rejects its own children. And yes, a huge part of replacement is about power and power dynamics.
1
6
105
Timo Horn makes more sense as a stop gap until Neuer gets better and is more likely to want to move in January.
4
The funny thing is that I was just browsing Youtube and I accidentally stumbled on this incredible Onion satirical video from 13 years ago(they used to have very good content). It seems to capture the essence of Google and the industry perfectly: piped.video/watch?v=lMChO0qN…
5
1,734
I remember when Yud was (rightfully) regarded as a niche cult leader. The fact that his ideology ended up so dominate in AI spaces is deeply concerning.
1
4
369
Are you aware that part of Israeli doctrine is enslaving Palestinian civilians and using them as human shields/human minesweepers? apnews.com/article/israel-pa…
1
5
85
Gemma 3 4b, which is a real 4b model which has been out for months, is only 10 points lower.
1
4
9,630