Deadlock Sock

Computer Scientist, I code sometimes ... Working as a Business Analyst, making my skills convincing business users it was their idea to begin with.

[Deadlock Sock @DeadlockSock]

@theo opus 5.5 is why I went back to anthropic the quality of code produced is much better. To be fair GPT will get anything you ask it done but at what cost? It's a deal with a devil there is always a price you are unaware of.

[Deadlock Sock @DeadlockSock]

@dhh What models do you use that you put so much faith in? The other day I caught GPT doing quadratic permission tracking which is probably fine for < a few 100 users but what does your workflow look like? I might be prompt min'ing

[Deadlock Sock @DeadlockSock]

@sama I sit of the fence of this because AI is inherently biased by employees and training data. The problem is I think being able to explore and ask questions to what effectively is earth's dictionary. Is the single most useful thing you can do for critical thinking.

[Deadlock Sock @DeadlockSock]

@wesbos My main question is even if this was 100% convincing is this what we really want as a society? I much prefer dot / muse over this...

[Deadlock Sock @DeadlockSock]

@stepbystepnomad @AlphaOM3G1 @cherry_mx_reds We are getting ripped off out here in NS lol I suppose the most important thing with self-hosting is really how much electricity costs where you live. My wife is from a country where it's about ~0.50 CAD kWh making the calculus much different Thanks for the numbers, it's helpful!

[Deadlock Sock @DeadlockSock]

@stepbystepnomad @AlphaOM3G1 @cherry_mx_reds Thats quite low, are you having electricity subsidized? In Nova Scotia I would pay 300 every two months which is about 150 a month. without running specialized hardware. what's your KW/hr price where you live? Never the less thats not bad at all.

[Deadlock Sock @DeadlockSock]

@stepbystepnomad @AlphaOM3G1 @cherry_mx_reds How much did your bill increase monthly in raw numbers? How much CAD?

[Deadlock Sock @DeadlockSock]

@theo Usage actually pretty good given the API cost increase. I can even get away with using it on fast mode for localized prompts. If people are running out of usage they might be holding it wrong nlg.

[Deadlock Sock @DeadlockSock]

@Akthamix @OpenAI Skill Issue

[Deadlock Sock @DeadlockSock]

@ThePrimeagen Safety first! Am I right fellas?

[Deadlock Sock @DeadlockSock]

@thsottiaux I like codex because it does exactly what I say, I hate codex because it does exactly what I say.

[Deadlock Sock @DeadlockSock]

@0x6080 @ThePrimeagen @bcherny how dare you insult harpoon. We got a real Moby-dick over here.

[Deadlock Sock @DeadlockSock]

@unclebobmartin @ThePrimeagen It sounds like I would have a better experience plucking my eye lashes out one at a time. ngl.. I refused to believe this is "fun" by any metric of the word.

[Deadlock Sock @DeadlockSock]

@sama We need goblin mode. I want it to go full Hob Goblin, with a little goblin jumping around my screen, picking up icons and throwing them.

[Deadlock Sock @DeadlockSock]

@sama If OpenAI could get a diffusion LLM to 5.3 codex level intelligence. Most people would be extremely surprised and probably exclusively use it.

[Deadlock Sock @DeadlockSock]

@sama The real concern is verifiable domains with ignorant management. E = MC^2 allowed for world ending weapons to be created. That was not intended by the creator. Count blessings and hope the majority of the world press the blue button.

[Deadlock Sock @DeadlockSock]

@sama This is only because humans are still very much needed in the loop. However, GPT 5.5 is extremely good, the issue is motivation. We will see real change when these models become more self driven. My openclaw still often needs me to say "yes go do that" because it was implicit.

[Deadlock Sock @DeadlockSock]

@thsottiaux @TheGeorgePu They got that much faith in their design tool huh? interesting.

[Deadlock Sock @DeadlockSock]

@SierraHotel28 @davepl1968 Instructions only happen in different orders when there are race conditions, I agree. Also if you use f32, unless you run a ~300B model you are unlikely to see the compounding errors materialize. However, yes quantization could potentially make even small models non-deterministic