Dave Lawrence

@lemire @pmarca You care about quality but do users? Yes, vibecode is slop, but people love slop. Look at viewcounts on slop videos Slop can be cranked out for niches quality can't afford to serve, it will crowd much of it out

In this video, @pmarca describes how some people have stopped sleeping because they code with AI. His model is that the ‘opportunity cost’ of sleeping is too high, meaning that they can produce so much value with AI that they’d better not sleep. My model of what is happening is a novelty effect mixed with gamification. There are people who are naturally drawn to new experiences and tend to embrace them too much. Also, building with AI is a bit of a game. There is a Minecraft feel to it: you quickly build a house and then walls around the house and so forth. I’ve seen it with some graduate students who produce massive amounts of ‘stuff’. Some of them become convinced that they are the new Einstein. They are not. What we have is a genuine glut of ‘research outputs’. We have been there before, repeatedly. For example, about 20 years ago, blogging became a ‘thing’. The WordPress empire came to be. Everyone and their dog had a blog. The lady running my university started a travel blog soon after she had to leave due to her general incompetence. I don’t expect that her career as a blogger lasted very long. I suspect that the same will apply with AI. Throwing stuff at AI is fun. Building valuable software is difficult. Even though we all have access to a word processor, few of us become novelists. AI will not turn most of us into software architects. What is happening is Cargo Cult Software development. Cargo cultism is a reference to Pacific Islanders who would copy the US military after the US was gone, in the hope of getting the goods that the US was providing. Software is not done for. We just have a new tool. Building useful software is still hard. Some people like it, but most will tire of it and find faster tricks. People will add skills or other lightweight techniques. There will not be a world where everyone writes their own operating system or browser.

@lowlandsapien Klemzig can’t claim “oldest surviving continuing German settlement” because for 18 years it was renamed to something less politically controversial: Gaza

Hahndorf is the "oldest surviving continuing German heritage" town but the actual oldest is Adelaides first suburb, Klemzig. My 4th great grandfather got a job pulling nails out of the ship and straightening them for the return journey, while the rest of the family walked from Port Adelaide to the row of tents which would be become Klemzig. It became majority British fairly quickly. (Queen Adelaide was also German)

@norvid_studies Have you seen the latest Napoleon movie? To me, was bizarrely focused on relationships at the expense of other things. Lots of men had wives and mistresses but how many redesign artillery corps and introduce standards and measures?

By his late forties, Napoleon started to suffer from numerous medical ailments, including kidney disease, bladder stones, chronic bladder and prostate infections, arthritis, gout, obesity, and the chronic effects of smoking. In 1856, Dr. Robert Ferguson, a consultant called from London, diagnosed a "nervous exhaustion" that had a "debilitating impact upon sexual ... performance"[185][page needed] which he also reported to the British government.[184][page needed]

Search for a wife Soon after becoming emperor, Napoleon III began searching for a wife to give him an heir.[citation needed] He was still attached to his companion Harriet Howard, who attended receptions at the Élysée Palace and traveled around France with him.[citation needed]

@nizzaneela To make a curve, so you can eyeball cost/accuracy tradeoffs, to design an experiment to actually see if it is true

@nizzaneela There is real life biological and technical randomness. Did 200 runs per point, probably ~40 points (altering seq depth vs number of variants)

When your Monte Carlo experiments use a random number, are you:

@techbromemes This has a bug: needs case insensitive match

@Rothmus No one would be against data centres if we used them to run the calculations on this

The goldilocks zone

@SRamirez68083 Prediction markets are tax on bullshit. Does she have a bet on "robots self replicate from rocks in early 2030s"?

The computer people are so detached from physical processes they end up with a level of thinking equivalent to “water comes from the tap”

20 mins of vibecoding a joke, and it got 400 GitHub stars in a few days, more than all my other code combined. Time to pivot from clinical bionformatics to shitpost driven development

Colleague says all cells ruptured, so they are dead already. Someone points out that it's possible they could be revived in the future. So seems not murder, but possible desecrating a corpse. Witness quotes me, I say it was just a co-incidence

Watching someone use liquid nitrogen at the lab today, and wondering the legal implications were they to accidentally freeze themselves, and I said "hasta la vista, baby" and shatter them

@michael_timbs Maybe they can get jobs in Adelaide

One of the interesting and recurring themes in research around tacit knowledge is that it is usually cheaper to keep this kind of work going because the next time we want to do something like this we'll spend more retraining everyone than we would've spent just having a never ending cycle of work

@akarlin > Programming is much more evolutionarily novel and has not been sexually selected for at all Ooof

I think there's a pretty simple and intuitive explanation for this. Good writers have been sexually selected for 5500 years, and writing and reading have always been the primary mode of symbolic analysis (solving problems through abstract thought). It is also loads heavily on far more ancient linguistics modules some of which like the FOXP2 gene we even share with parrots (a notably verbally tilted species). Programming is much more evolutionarily novel and has not been sexually selected for at all. On some universally neutral scale of cognition, the top human writers are several S.D.'s better than the top human coders. Consequently, whereas AI has already blasted through the top tiers of human programming ability, reaching the peaks of human writing ability is still probably a couple of years away.

@SashaGusevPosts @AndyMasley If you went down a bad path in plan, delete it. It is generally better to try again from a different prompt to get it right rather than correct an error as misunderstanding sticks And give examplees of what to do, not "don't think of an elephant"

@SashaGusevPosts @AndyMasley They get more distracted as context fills and am constantly manually managing context size. I get it to write a plan, review and refine via conversation. Then ask for a prompt to implement plan after clearing context. Copy prompt, /clear context, paste to implement plan

For people who are extremely skeptical of AI being able to automate AI R&D, what specifically is the magic spark that humans have that AI will never have that's relevant to AI research? If you imagine the typical AI researcher today, what knowledge or abilities are in their brain that AI will never approach? If AI achieved those abilities, we could effectively hire gigantic numbers of AI researchers, way more than the human researchers we have access to. Maybe you disagree with the timeline, or you believe the bottlenecks in AI research are elsewhere, but I find that a lot of people just treat this scenario as sci fi because it "feels" crazy, but don't often have a solid reason it couldn't happen.

@SashaGusevPosts @AndyMasley Same thing with many humans. Work with the agent to create an overall plan, decompose the problem into subtasks (of less than a human day's work) then have it do separate plan and implementation steps for each subtask

@AndyMasley A major limitation of LLMs right now is they are unable to prioritize goals in open-ended tasks. They will routinely go down pointless side-quests while ignoring critical errors. Over the course of a series of iterative decisions these errors explode.

Downloading pdfs for Claude, clicking "I am not a robot" so the robot can read them

@norvid_studies Watching other people use a computer

what else is like this