@irl_danB Why can they delete their db via an api call is the real question
English
barry farkus
955 posts
barry farkus
@paimon2cool
circling something deep
cyberia Katılım Mart 2024
109 Takip Edilen79 Takipçiler
this post is sad and enlightening
just an absolutely wild level of misunderstanding about these systems, even from technical people using them in production
I’m at a loss for how to remedy this gap as urgently and as quickly as needed
JER@lifeof_jer
English
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.
English
@gabriel1 The optimal number of hours isn’t a fixed number. Depends on the week
English
My biggest challenge with vibe coding / agentic engineering lately has been getting stuck in what I call a "plan doom loop"
- have AI write a plan
- review myself, seems good
- have AI review plan, it always finds something
- repeat
It drains my time and energy to determine how important the "findings" really are
Who has solved this
English
@yacineMTB Weren’t you in Costco building a robot from your phone a few months ago?
English
@k1rallik @Soul0Engineer What’s the difference between this and squeezellm published ~2 years ago?
English
> been paying $200/month for cloud AI APIs
> laptop: M2 MacBook, 16GB RAM
> tried running models locally, garbage quality after 4K tokens
> read this TurboQuant breakdown on Tuesday
> applied 3-bit KV cache compression
> same MacBook now runs 100K token conversations
> quality: identical to cloud
> cancelled all API subscriptions Wednesday
> it's been 3 days
> saved $200/month forever
> with a free algorithm from a free paper
> my MacBook didn't change. the math did
BuBBliK@k1rallik
English
Full quote.
“Or do you think it just fixes itself by doing this for long enough? I assure you that is not the case.
Again, I agree hallucination rate tends to be lower because more tokens and good post training make self correction more likely”
English
“do you think it just fixes itself by doing this for long enough? I assure you that is not the case.”
I can’t tell if you’re fucking with me or just delusional, or maybe you think someone reading this would be too lazy to scroll up to clearly read you saying the opposite?? 😂😂😂😂😂😂 done with you child
English
I literally said reasoning gives them more chance to self correct, but that doesn’t prevent hallucinations. Scroll up.
Arguing with retards is a bad habit of mine so going to stop here, but ask yourself why one of the big labs hasn’t solved a millennium prize or some other unsolved problem? If it was as simple as just letting reasoning continue for long enough it should be easy right?
English
Nah bro you said they couldn’t even sell correct 😂 you said more reasoning wouldn’t result in better answers. Don’t lie. Yeah I said in theory, as one very small part of my answer. There are models that can run for days, OpenAI has talked about this, and they appear pretty flawless, but they aren’t for public use, obviously that’s way too expensive and computer intensive to offer as a service, but they exist, because, say it with me, “more reasoning - better answers.” Who knows what the limit is. You were arguing against this stop lying, anyone can scroll up lol
English
I’ve read the deepseek paper many times. Sure, reasoning tokens improve output - I’m not disagreeing with that. Your claim was that “in theory it could go on forever” - the passages you’re quoting don’t back this up like you seem to think. No one who actually knows how LLMs work is claiming this.
English
😂😂😂😂😂😂 okay this is when deepseek learned that more time equals better answers. The paper even shows this example of it correcting itself thanks to more tokens, something you claimed it wouldn’t do. ALL of the models show this trend. More time - more accurate answers. Sucks to suck
English
@aifilmmaker @Mechanic76Jack @truthache68 I know you’re making it up. Show me the paper you think makes this claim.
English
It is true lol you think I’m making it up? I’ve read the papers on these models, the chart is always there. In theory it could go on forever, but we don’t have unlimited compute.
IMO AGI already came and got went anyway. It just means general, as in not one specific task like chess. These models can do chess, math, poetry, etc all in one model. Thats AGi but people keep changing the goalposts lol
English
English
It literally fixes itself by doing it for long enough, on a graph it’s like a straight line going up - the longer the reasoning, the more accurate the answer. The only limit here (that we know of) is how long you are willing to let it run. I can find you a chart that shows this lol
English
So what if it hallucinates within the … blocks? Or do you think it just fixes itself by doing this for long enough? I assure you that is not the case.
Again, I agree hallucination rate tends to be lower because more tokens and good post training make self correction more likely. But reasoning models in no way prevent hallucinations.
imo, hallucinations are a good and necessary part of models anyway.
English
That’s not how it works. The reasoning is part of the output, and can go on for as long as you want, hopefully ending up on the right answer eventually but it’s not usually the first thing - the longer the reasoning the better the answer, that’s how they have models winning math and coding competitions, you just let them go for hours on end. It absolutely prevents them in a meaningful way, to the point where they basically don’t exist, especially for the long long reasoners they have (that aren’t publicly released)
English
@FrankiePu1itzer @aifilmmaker @Mechanic76Jack @truthache68 Yeah and that’s a different thing.
There are also transformers that predict blocks of tokens at once. But again, different thing.
arxiv.org/pdf/2404.19737
English
Right but it does literally spit out the “first thing”. But the first thing is just conditioned on useful reasoning traces. I agree it reduces hallucinations in practice , but it doesn’t actually prevent them in a meaningful way. I’m disagreeing with your take - not agreeing with the video
English
It absolutely changes it lol, but SIGNIFICANTLY giving them a better chance, hallucinations are so rare now that it’s hardly a problem anymore, if you’re using a SOTA model. It’s not as simple as just spitting something out even if they don’t know, like the video claims. There are tons of safeguards. Sure sometimes they fail but it’s not like the video says, sorry.
English
@aifilmmaker @Mechanic76Jack @truthache68 Reasoning models don’t change this bro. They still predict one token at a time. Reasoning traces just give them a better chance of those predictions being accurate
English
@Mechanic76Jack @truthache68 If LLMs just spit out the first thing, then sure. But they haven’t done that in years.
English
@JadeCole2112 When was the last time you wrote code professionally?
English
Jimmy Carr drops a wild take on AI + physics
"Strip the screens away, and we're still living in the 1970s. Nothing real has happened in physics since 1972. String theory? Nowhere.
But point AI at real physics now? Everything else is just stamp collecting. Physics gave us all our technology.
We could get a 50x productivity leap and true human flourishing... or it could go the other way."
The real revolution isn't just AI—it's AI meeting physics.
1:02 clip inside—mind-blowing perspective.
English
if LLMs will make software engineering obsolete, why don't they just generate binaries instead?
not code & compile under the hood. Object code is data too.
why do LLMs have to generate things that humans create?
English
@Midnight_Captl @dwarkesh_sp @karpathy Karpathy is just indecisive. Flips position in this every other month.
English
The change from the @dwarkesh_sp podcast 2 months ago vs. the tweet below from @karpathy is genuinely insane
It’s a night and day difference. We went from “these models are slop and we’re 10 years away” to “I’ve never felt more behind and I could be 10x more powerful”
This all changed with Opus 4.5. It will be looked back on as a historical milestone.
Andrej Karpathy@karpathy
I've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become available over the last ~year and a failure to claim the boost feels decidedly like skill issue. There's a new programmable layer of abstraction to master (in addition to the usual layers below) involving agents, subagents, their prompts, contexts, memory, modes, permissions, tools, plugins, skills, hooks, MCP, LSP, slash commands, workflows, IDE integrations, and a need to build an all-encompassing mental model for strengths and pitfalls of fundamentally stochastic, fallible, unintelligible and changing entities suddenly intermingled with what used to be good old fashioned engineering. Clearly some powerful alien tool was handed around except it comes with no manual and everyone has to figure out how to hold it and operate it, while the resulting magnitude 9 earthquake is rocking the profession. Roll up your sleeves to not fall behind.
English
We've got big plans to improve the core ChatGPT experience in 2026. What are some thing you'd love to see?
Even small ideas welcome! Looking forward to getting them built 🙏
English