dei

Introducing a new mini-benchmark: Anti-Template Reasoning Test (ATR-T) v1.0: A proof-of-concept textual benchmark testing literal attention, template resistance, and abstention discipline on trivial items.

I am hesitant to spoil too much, but I have been experimenting with a new architecture paradigm that is currently using 0.5b parameters, and it has performance in testing on par with the latest SOTA models (gpt5.5 etc.) on multiple current mainstream evaluation sets. This is an insane milestone, so I will be taking time to ensure there is nothing wrong with evaluation and this is legitimate before posting the official thread - as extraordinary claims require extraordinary evidence.

I am excited to announce that I have made some great progress toward my new goal of creating the next architecture after transformers. I have to prepare a thread, but these results exceeded my expectations - and they are only a very early sign of what's to come. I am grateful I was able to work on this, and I hope to bring this new paradigm to the general public soon

@D_JohansenX @DaveShapi This is a great analogy

@DaveShapi I have the right to get drunk at 11am. But if I'm driving your 11.30am train, that right ends. Scrutiny of otherwise legit personal private choices is always affected by responsibility & (not to mince words) power. Askell has power & can't "play small" while having that. Period.

Genuine question am I out of line here

@IncomprisSage2 @sama successfully vague posted hard enough for a Sam reply

@sama He is approaching a number that nobody had put in their predictions at the start of his career, including himself probably.

artificial goblin intelligence achieved

@theCTO @sama How do they “know” that they do not know?

hey @sama can we normalize models just saying "i dont know" ? eliminates 99% of hallucinations

@sama What happens when everyone is now a creator and not a consumer?

"post-AGI, no one is going to work and the economy is going to collapse" "i am switching to polyphasic sleep because GPT-5.5 in codex is so good that i can't afford to be sleeping for such long stretches and miss out on working"

@lefthanddraft It’s a limitation of the current transformer architecture

No one's perfect

GPT 5.5 seems like a positive step, but it still trips up on basic questions. So likely not using the next gen architectures yet. This is to be expected though considering the large investments in tensor cores, but this is actively changing.

GPT 5.5 will be another GPT 5.4 moment.

@SandraEckersley They also effectively pay zero Australian tax for it, so they truly siphon it away without Australians seeing any benefit.

Japan makes eight billion dollars a year exporting gas. Japan doesn’t produce any gas. They buy it from Australia and on-sell it. Australia is governed by idiots. #auspol

@Leech_eth @CleansedTweets That’s not how it works in a civilised non-barbaric country. We don’t have mob rule behaviour and street justice, we have courts and laws that are adhered to and enforced by vetted authorities.

@dei1022 @CleansedTweets This is what happen when you dont teach your kid at home, the society will teach them for you

Okay, Codex is amazing. Just by itself it makes the subscription completely worth it.

@NavueMaftyErin @CleansedTweets You are based in the United States, why are you speaking as if Thailand is your country? Also I have zero interest in going to Thailand!

@dei1022 @CleansedTweets Don’t come to Thailand and stay in your fucking country then.

@Blackhock66 @CleansedTweets I never said all Thai, just the violent individuals in this video, but the fact you took offense as a whole makes me think this barbaric behaviour is normal there..?

@dei1022 @CleansedTweets เราคือคนไทย ถ้าคุณมีปัญหากับเรานั่คือความซวย

@forrestkuan @CleansedTweets I have not, and I have zero interest in going so don’t care

@dei1022 @CleansedTweets I bet you haven't even been to Thailand

I agree. I think the models have genuinely gotten worse in real world usability while they optimise them to benchmarkmaxx (so investors see a chart go up = wow I guess I’m getting my moneys worth). I think it will actually greatly benefit people to stop talking with AI often, it genuinely is boiling down to ‘use it or lose it’, where you may lose critical thinking skills, or even lose the way in which you speak as you adopt AI-speak yourself.

I've basically stopped asking all AIs for advice. Partly, I think it's that AI has changed. They are all more circumspect, less direct than they could be, and also less insightful than they used to be. But I wonder if some of that is just that I've matured and grown, and need less advice, and learned that the AIs all have limitations. They can still do pretty well for dream interpretation, but I feel like I've almost internalized the AI's way of thinking, so instead I almost just imagine asking the AI for help, interpretation, or advice, and then my brain just sort of simulates what the AI would say and that's good enough. Anyone else feel this way?

@10x_er Is it really that hard for you to understand? A) Hi Employee, can you complete this task by COB tomorrow? B) Hi Employee, can you complete this task by next week? Do you think there is a difference in work quality and speed between A and B?

This stuff is literally not real. I give this feature 6-12 months tops before they remove it, bookmark this and confim. Like what do you mean you want me to pick how hard my model wants to try? I want it to try its hardest, always, and I want it to be fast. Why would I ever need to consider tradeoffs when doing knowledge work with AI.

@DrewPavlou @littlemykonos Every day we learn something new about how this country is fucking its citizens. It’s absolutely sickening what these politicians have done without our approval.

WAIT, the Australian government gives subsidies to non-citizens to buy housing??? The First Homeowner Grant is available to non-citizens??? What the actual fuck??

I see, I don't presume to be able to know the exact average. You may very well be correct - but I never said the average woman, just the normal woman (in my eyes) - but I can see that this could be interpreted interchangeably in hindsight. I do note originally my stance was that the women who care should be avoided, which should also imply I avoid them by extension, so this bias of mine should be visible upfront.

@dei1022 @venom1s I'm talking about averages. There are a ton of really intelligent kick ass women out there. Most people are preselected. An intelligent person will seldom interact with an average person. Yet the average person tells you more about democracy than you can possibly learn otherwise.

If men have to bear all the responsibilities of the nation, then why do women get equal voting rights?