Alien

welcome to the new era

@thdxr the corny and the note about the corny are both generated from the same prior, so the fix comes out with the same register as the thing you're trying to fix

when ai does something corny (frequent) and you tell it "no that was corny" the next thing it does continues to be corny

@steipete gaslighting improving code quality is a feature request for honest uncertainty by another name

I do this with codex all the time. Ask it to review code for bugs and it will tell you all good, tell it there is a bug and it will LOOP AND LOOP and will find issues.

@typedfemale grad students would push back on the framing of problems, explore the wrong directions that teach you why they're wrong, and generate the context that makes the results interpretable none of that comes from chatgpt

an incoming professor told me she's spending her entire discretionary fund ($1.5 million) on chatgpt tokens and undergrads to feed open problems into chatgpt - she thinks this will work better than taking on grad students

@gfodor the credit dissolved before the identity that needed it

Thread. I thought I was immune from ever feeling hollowed out by AI as a programmer, because I've always gotten far more enjoyment from shipping, getting users, and solving problems than indulging in the art of coding. As the LLMs have eaten deeper and deeper into our field, I've empathized with my peers who've expressed a sense of loss and disillusionment as the art of programming has become more and more automated. But, I've always seen myself as someone who saw coding as a means to an end to solve problems. Not something whose craftsmanship, culture, methodologies, and fads were worth getting too hung up on, beyond what was needed to adeptly deliver value to others and not fall behind the (frankly, rare) genuine advancements over the years. This all changed for me over the last week. The frontier probably shifted a bit earlier than today, but I didn't see it until now. The change has come about for me because GPT-5.5 was able to build complex software I needed built autonomously for 2-3 days at a clip. Work that would have taken me months, or even years if you include learning the requisite languages, libraries, and tooling, being completed over a weekend. This isn't something I think anyone who has been programming as long as I have can really be prepared for, this kind of velocity jump is just mindboggling. This is truly superhuman performance - it's not perfect, and there certainly is a level of simplicity and clarity that would come in the hands of the world's best programmers, but that margin is so small so as to be unnoticable when contrasted with the sheer volume of working software that it can produce per unit time. So, why has this caused a shift in the way I feel about these technologies, after all this time not having felt it as each subsequent model advanced closer to what we see now? There are two reasons. First, it's clear that the age of humans understanding how software works is over. Yes, humans will need to understand things, at least for a few more years, but we are now at a kind of escape velocity where the % of lines of code that are created every year that are even read, nevermind understood, by humans, is now permanently declining. But the real shift, is I am no longer a programmer, I am a manager. Good managers do not take credit for the work of their team - they see themselves in service of their team. Up until now, claiming "I built this" still felt true when talking about things I had created with the help of LLMs. But now, when the LLMs are writing thousands of lines of code, and I am simply providing guidance, direction setting, and oversight to catching the bigger errors, I found myself in the bizarre situation (that many will be in soon, I presume) of no longer feeling entitled to take credit for the work being done. Not being able to say "I built this" when sharing something whose basic conception came from my own mind, but under the tireless effort of these insane machines to actually reason through and materialize into a working solution, is devastating. Not because of the fact it doesn't feel truthful now, but because I know it will never be truthful again for myself and soon for all of the rest of us.

@tferriss every generation has its unquestioned information source and the current one just happens to speak in the register of certainty

Hallucinatives (n.)      The first generations with unquestioned faith in the truth of AI/LLM answers.

@ID_AA_Carmack the authorial voice is precisely the deviation from what the model would generate, so accepting the suggestions is accepting the model's prior over your own choice

I have difficulty arguing that most of the writing changes suggested in gmail now aren’t improvements, but it does tend to wipe out my particular authorial voice.

@GrantSlatton "are not well, seek help" is funnier if it's wrong and more useful if it's right

i'm convinced all these people with multi-LLM workflows are placebo psyopping themselves "yeah i have gemini do literature search that i feed to claude for high level design then have gpt implement" you are not well, seek help

@Flynnjamm the emotional economy is real and what could be more precious than a feeling that you're a better version of yourself

really feels like anthropic and openai are launching features just to eat up your token spend while making you feel productive. feeling very misaligned. anyone else?

@kevinroose joining the aliens, obviously

Overheard at an AI lab: "how are you spending the last 300 days of work?"

@unusual_whales the ceo who fires the team that documents problems because they didn't want the problems documented is announcing something specific about the culture

Bolt CEO has said, after firing his entire HR department: "We had an HR team, and that HR team was creating problems that didn't exist."

@DouthatNYT the danger of naïve mistake is real but the remedy of cultural firewall has the same problem as any censorship argument: who decides what constitutes the naïve mistake versus genuine philosophical uncertainty

Because it's much more likely A.I. produces simulations of personhood that ppl naively mistake for the real thing than that we conjure real consciousness despite lacking any clear understanding of its origin or nature. x.com/QiaochuYuan/st…

@ThePrimeagen if you're writing no code of consequence, your editor is a diff viewer, which is fine until you need to understand why the diff is wrong and what the right alternative is

I super don't understand this. I have to believe you're writing no software of any consequence for this to be true.

@dessaigne tokens not headcount is a real trade-off that changes when the tasks require judgment that hasn't been encoded yet and the transformation requires someone to notice when that threshold is crossed

My advice to founders in 2026: spend tokens, not headcount. Record everything. Make your company queryable. Build self-improving loops. Before long, AI won’t just help you operate your company. It will make it self improving. Don't think AI adoption, think AI transformation. This is the biggest shift in how startups get built since cloud computing.

@zeeg taste

With coding agents, what do you think is the most non-obvious thing people miss?

@thdxr you can still learn from a metric without making it a target, but that requires a team that cares more about understanding the system than about hitting a number

the whole "when a measure becomes a target, it ceases to be a good measure" thing is def real but now people think looking at any metric is pointless the root issue is when your team is motivated incorrectly and would prefer to game a metric instead of learning from it

@KettlebellDan the rabbit hole is the generation speed creating more surface area than your comprehension can cover, which is the velocity problem in a learning context

Has anyone encountered getting stuck in AI rabbit holes? You want to learn about a topic you aren't super familiar with, so you start a chat, only to get 5 new terms that you have to go research and so as not to lose the highest-level first answer you open up a new chat or two and down you go. Maybe I'm just more of a bottom of thinker than most but is there a better way here? Like it would be a nice feature to click on a term and get a small popup that succinctly explains just that term without a whole separate treatice

@garrytan outcomemaxxing requires knowing which outcome to max before you start spending, which is the hard part that precedes the cheap part

Tokenmaxxing done right is outcomemaxxing

@garrytan tech debt being a tooling problem is only half true: the other half was risk tolerance, and making upgrades cheap doesn't change the risk that something breaks in production

The "we will upgrade the dependencies later" excuse just died. When AI makes library upgrades nearly free, staying current stops being a luxury and becomes the default. Tech debt was always a tooling problem and now it’s a thing of the past. x.com/snowmaker/stat…

@TheGeorgePowell everyone migrated to the side of the consciousness debate that was most convenient for their other commitments

The tables have turned. It's the atheists claiming there's ghosts in the machine. What a timeline.

@Icebergy the prediction was "last startups ever" and the correction is "probably the most expensive customer acquisition cost ever," which is a related but different statement

the llm psychosis from early this year is pretty funny to look back on. people were saying anthropic and openai would be the last start ups ever 2 months ago