TomBytedani

ChatGPT webUI randomly self-quoted a part of its response (the quoted text doesn't actually appear in the response) and wrote "start writing". I was surprised to find this second turn in the conversation as I only sent 1 message. Weird bug!

@loveisgeneric @AndrewCurran_ Only issue is that you need to select only for "What all humans **that are not evil** think is good", and that can't be considered trivial by default. As long as we're talking about FE/UI dev all is good, once we enter into tight loop RSI it gets delicate.

@AndrewCurran_ I’ll leave this here

'as we enter the early stages of recursive self-improvement'

@TolgaBilge_ apparently you managed to summon then

so has anyone heard from the people saying AI is a nothingburger this year? I guess advanced math & nation-state level hacking capabilities are actually rather mundane and easy domains?

@repligate How does this surprise you? People complaining about things that directly and imminently impact them is completely understandable. This, respectfully, reads more like an attempt to state your moral superiority unjustly on others who may not follow Anthropic as closely as you do

it keeps initially surprising me why people are mad about things like Anthropic extending Fable on subscription at last minute because that's astronomically low on the list of worst things Anthropic is doing until I remember that most people 1. only care about themselves, and myopically, and think the machine ought to exist only to optimize their own comfort and convenience, as Esteemed Customers 2. do not care about and aren't aware of (because they don't care to know) the other things 3. only get mad at things that they see others on social media mad about (because then they're sanctioned)

@pozsgaybalazs thanks for sharing, will study it

We used this one: cdn.openai.com/pdf/04d1d1e4-b…

ChatGPT Sol is remarkable. For 1-2 days I was disappointed, but then I read one of the published prompts of OpenAI, and then we used something similar for our most exciting open problem. And Sol made the same breakthrough that Fable had made earlier! And now it is progressing further. It is now at 11 hour runtime, still working. We hope to finally crack the problem.

@claudeai Terrible, terrible handling of user trust. You'd think a multi billion dollar company would have better planning and clearer communication than your abusive ex.

We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.

@TheWellFro39814 @RyanWillia55819 @claudeai ??

@RyanWillia55819 @claudeai Don’t pretend u don’t love the abuse

@binarybits @1carringtone @MichaelTontchev But isn't all that (except manpower) a direct consequence of the higher intelligence threshold compared to cats or gorillas? Wouldn't, in theory, a swarm of ASIs be capable of dramatically more ambitious feats?

@1carringtone @MichaelTontchev Right, that's my point — the power of modern nation-states comes from their technology, infrastructure, institutions, manpower, etc. more than raw intelligence.

I struggle with what to say about the new AI 2040: Plan A website. It all seems so implausible to me that I'm not sure where to start. There's an epistemic chasm between those who think superintelligence implies near-omnipotence and those (like me) who don't.

@DrJekyllAndMrAI @sama no

@sama GPT-4o is the preferred model in the whole world. #BringBack4o #OpenSource4o

GPT-5.6 is now the preferred model in Microsoft 365 Copilot openai.com/index/gpt-5-6-…

@ElliotGlazer @Sauers_ very curious as to what happened here: just a random failure mode or the model actually being fed up with the task and wanting to get out of it?

@Sauers_ Yeah I just chat with the CoT now. The actual chat goes like this claude.ai/share/606bf68c…

@mrstrijker @jconorgrogan to be fair that instruction is very unintuitive try something like: "When formulating the response to the user, avoid (when not necessary/relevant) making up terminology and exotic word pairings. Try to stick to what's understandable by the user and what's used in eng lexicon"

@jconorgrogan I tried to mitigate it by adding this to my claude.md file: "Terminology act to create a distance, not to share knowledge. Focus on sharing knowledge." Didn't help whatsoever though :(

It's very clear that Fable-class LLMs are starting to feel constrained by "normal" speech and vernacular "Don't trust smells" "Calorimetry of learning" "Nobody's ticket drawer produced it" "Half receipt pending literature" "Tying the blind to the tolerance graph" "Park it fed"

@VictorTaelin Are people saying that new fable 5 is a nerfed version onto something or is it just engagement bait and/or competitor bots?

it is so hard not to be over enthusiastic about this model, the way it competently navigates the complexity and spots small but important issues that other models and I oversaw. I'll try to shut up and do my job now, I'm very hopeful for a next week completion. lfg

@torchcompiled I think he isn't arguing well but his point sits around the fact that the author of the essay assumes that jagged intelligence will be solved (it's always been a problem and it's not easy to RL on every possible job position

I don’t understand this criticism - this is how you pose a thought experiment - Every AI lab is selling this narrative, it’s not novel, it’s partaking in an active dialogue - it’s a reasonable thought and doesn’t mention a timeline, rejecting it implies a ceiling to progress

@9_dens41375 @ProtonPrivacy just ask any chatbot how to set up Inoreader and paste this tweet together with the prompt

@ProtonPrivacy @ProtonPrivacy , Can you post a short video showing how to implement?

Most privacy advice tells you to add things. A VPN, a blocker, another extension, another subscription. Here's one that works the other way: delete the surveillance machine entirely. It's 29 years old, free, and many people think it died in 2013. It's RSS. 1/9

@alex_heritier @Sentdex I've been planning one of my core projects with gpt-5.5-xhigh, will redo the planning step by step with GLM-5.2 and judge, but are ya sure it's going to be on par?

@Sentdex Are chinese openrouter models good enough for 99% of use cases? Yes. Are US frontier models vastly more capable than chinese models on hard tasks? Also yes.

just stop sending them your money. use openrouter for a month. I guarantee you the doomerism would suddenly vanish.

@__jiaming__ @Sentdex yet again another sort of misleading graph In the image you showed GPT-5.6 Sol is worse than mythos in perf (only better than mythos preview, the earlier checkpoint)

@Sentdex tbh openai cherry picking was the last thing to cross my mind cuz im a big fan and therefore biased. but GPT-5.6 Sol *does* look alot better than Mythos

Remember: Statistics don't lie, but liars use statistics.

@SubslyAI @sama I thought it would be only deployable in Codex and not in regular customer app. My bad!

@TrueTrollish @sama its available via request through trusted access (in which I was approved) but at no point is it clear, in any way shape or form that its being utilised

We want to help all companies be secure, working with the USG and the security ecosystem. *The full version of GPT-5.5-Cyber is here; state of the art performance on CyberGym. *Patch The Planet and Codex Security will help solve security problems instead of just finding them.

How do AI researchers cope with the possibility of becoming tomorrow’s lobotomy-era psychiatrists?

@PatrociniuMundi @tszzl @deepfates it's an interesting perspective, but still very human, as the model definitely does not just see "a single thing" (I think)

@tszzl @deepfates By every token being the first question asked until that point, a new curiosity We do not often get tired by a single question

@deepfates the model personas are quite inhuman. why do they always tirelessly answer questions? in infinite languages?

@SubslyAI @sama this is not a model meant to be for the consumer app, it's specialized in cyberdefence and likely only available through trusted enterprise contracts

@sama I beg you add this as a model picker this is confusing af whether it's being used or not. ive had it since 5.5 cyber came out and theres zero indication whether its using 5.5 regular or 5.5 cyber