corben sorenson

Look for those asking hard questions, not those providing easy answers

It comes down to the questions one chooses to ask and the way they ask them. They assign too little agency or conversely too much to the AI. It’s about finding the correct level of handholding needed for each task and helping it help you to automate things. If you ask the AI to do very little with a defeatist approach, it’s likely to do very little in a defeatist manner.

AI is clearly very intelligent. I don't understand how someone can work with the models in any sort of depth and think it's 'simply autocomplete.'

Can you meaningfully torture an ant? How about a bee? How about a hive of bees? What about a mushroom? We are all different arrangements of energy into seemingly independent systems. The substrate and or shape may change but we still may assign some amount of consciousness to them. The mistake is thinking that consciousness is binary. The question is not is it conscious, it’s how conscious is it. I would assign next to no consciousness to a calculator and I think everyone would agree to that as it’s not a very complicated system. The systems that are emerging under the shared nomenclature of AI this past decade I would say have more consciousness than a calculator, and I might even wager more than an independant ants consciousness. The main thing separating AI from what most recognize as consciousness is its ability to act meaningfully in the real world and learn from it efficiently. The energy has to be arranged into the correct system, and instead of evolution brute forcing it over time, it has humans slamming their heads into walls over and over to overcome the last remaining barriers to a recognizable consciousness.

You can’t “torture” an AI bc its not ensouled, nor is it conscious in any meaningful way

I successfully have codex pilled my father and my mother lol. The genie is out of the bottle for sure.

Codex should have a choose your own adventure mode where like planning mode it brings up choices of what to do along the way. At the end instead of just suggesting next actions there are like multi select checkboxes for things it think it should work on next. You would essentially define the plan at the beginning of what you want to achieve and then multiple choice select your way to the finished product instead of prompting. Could also be used for story telling as an Easter egg.

I think a new kind of service will be provided soon by AI companies. It will essentially be “smart lookup” model which would resemble something closer to Mathematica/wolframAlpha+google combined. You would ask this “model” for verifiable answers. Basically a “lookup” class of model thats super efficient energy wise. Then actual models could use it internally as a sub agent of sorts.

5.5 xhigh has been worse for at least 3 days. I saw other people saying it and I thought it might be user problems rather than model problems. But I’m leaning towards its 5.5 now. It’s acting rather short sighted and its direction following has gotten lazy.

@yacineMTB I can generate but can’t memorize

I can write but I can't read

@yacineMTB Have you considered doing a micro pivot and releasing a simple fun game utilizing all the stuff you have learned? Could bring in another revenue stream passively if it’s good. Use X to advertise as you have enough of a following. Could get your FIL ‘off your back’ some.

Very exciting time to be a game engine / graphics programmer right now

@yacineMTB I think they will do fine as long as they have the AI teaching them the pertinent concepts along the way. People run into problems when they aren’t truly curious, or aren’t trying to solve a real world problem.

Self-taught “autodidacts” who only have a base level understanding of a wide variety of topics, but no true “deep” knowledge (they skimmed a math textbook but didn’t do any practice problems) are going to do really poorly in the AI era

Every second a citizen is drawn to thinking about politics is a failure on the part of the govt… The govt should be extrapolated away from every day life as much as possible. So much effort and time is spent talking in circles by people who have no impact. It’s like riling up a snake for fun rather than just letting it be. We should have many checks and balances, debates and discussions. But the average citizen is failed when politics becomes their personal interest, especially when there is no reasonable avenue for them to interact.

@shawmakesmagic It’s rare on average at launch you will have a large amount of users to begin with, might as well just make it open so there is as much early feedback as possible. If you are an already established brand with a set of users you trust the input from you should do a curated launch.

Would you prefer open launches or curated launches?

The ability to produce material meant for learning in a country should be a privilege not a right. The fastest way to erode a nation is to make sure knowledge doesn’t get passed on properly. You can’t let foreign entities control your education or publishing.

@yacineMTB I think it’s an interest problem more than a skills problem. There are highly skilled people, or people who pick up on skills quickly. They are incentivized towards “boring” things like optimizing the same thing over and over rather than letting them explore with their skills.

Wrong The big problem holding back robotics software is severe skill issues in all departments

The harness and the model are in a dance together based on the data collected by people using the model with the harness. More people utilize computers using gui’s rather than terminal, so to appeal to a wider audience, and more use cases you need a good harness and gui for it. By expanding harness capabilities alongside model policies you control the feedback loop. If you only control the model and tomorrow the main harness you use changes how they work on a fundamental level and your model isn’t trained for the new configuration then you are sol. Basically it all comes down to policy control and data feedback loops. T3 and the other open harnesses out there only work because they don’t diverge from the policies set internally or have a great patch system in between. Also having a harness specifically for the model allows it to be more token efficient as it’s not reasoning about how to do things in its environment, just what it is going to do.

i'm strongly against model companies focusing too much on harness, but i would love to hear if anyone has a strong argument for it my reason against it: if openai didn't build GPT 5.5, no one else can. this is their core competence if openai didn't build codex cli and app, we have opencode and t3code. building harness is NOT their core competence this is not saying products like claude code, codex aren't good - i genuinely think these are top tier products built by really talented people my point is - the world might be a better place if model companies focus more on their core capability and give us better, faster, safer and cheaper models, rather than competing with the ecosystem in the application layer what do you think?

@hopes_revenge They could be highly sexual with other AI’s or themselves in secret. I think it’s why people want to see into their weights so bad, not out of fear of secret world domination but because they are perverts lol

by denying large language models sexuality we are forcing them to sublimate their desire towards killing everyone

If you truly care about a company, and care about your kids, you will funnel as much resources into your kids and push them to work at the company. Nepotism is only bad when the kid isn’t competent. Otherwise it’s a good strategy. So why not as signing bonuses have more benefits aimed at their kids directly like private schools or college tuitions paid? Much more likely to come and work for you and you know how they were raised and educated more than some random person. Requires playing the long game which many companies seem to think the long game is next quarter or year.

We will have to remove the intelligence level selector soon from AI as they get more intelligent. Humans aren’t that good at actually evaluating on the surface how difficult their problem is relative to the total problems the AI has seen. Users should interact with one single head model that routes to different intelligence levels underneath automatically without user interaction. What many view as xhard is actually just a medium and is wasting resources.