sashank_1509

In the past this book was interesting since getting good at it did somewhat help you become a better programmer. At the very least, your typing and ability to write a first implementation with fewer bugs improved as you practiced competitive programming, to some extent even your ability to problem solve and come up with good architectures improved.

That’s all out of the window now. If we are really in “do we even read the code?” territory, then this is just a hobby to do for fun, and doesn’t matter for any real world work. But even if we read the code, but most code is written by prompting and iterating with agents, I don’t know if this will be helpful at all.

Empirically I just asked Astra without using the internet to exactly reproduce the first paragraph from Hamlet, Othello, Pride and Prejudice, Sartor Resartus by Thomas Carlyle), Sybil by Benjamin Disraeli (1845). GPT-6 got everything correct: https://chatgpt.com/share/6ac2a72c-e8bc-83e8-8f42-da8c7542de...

No human on earth can copy at this scale. Yes AI is beyond a database lookup, it does have reasoning on top of this knowledge, but my original claim that if there is a spectrum between copying with minimal changes and creative inspiration with minimal copying, AI is one the copying end while humans are on the creative end. Humans are very bad at reproducing anything verbatim, even if they wrote it like a week ago.

I don’t think we should hold humans and models to the same standards. Humans have a very small working memory. Most humans cannot reproduce code they wrote even a year back exactly.

A model can reproduce large swaths of its training data exactly. It’s a different algorithm that powers its learning process (it’s why it needs trillions of tokens to even learn basics of language).

If there was a spectrum from copying on one end to creative production inspired from something else on the other end, the human generally lies heavily on the right end, while the model is much more on the left, that gap is large enough, that yes the model is in some sense “copying”.

Pretty sure no one on HN says AI dumb

Yeah, I think I would avoid video games with AI art or code, I only play indie interesting video games and looking at the public’s reaction I think they too feel the same way, so I do expect a smaller market for handmade video games.

Outside of that it’s hard to envision handmade anything in software in this everything mass produced cheap consumerist society.

He makes the exact opposite claim in the video. He thinks 20% is far too high and no one will pay for it once these tools are no longer available for free. He specifically cites the company Coverity that also found many useful bugs with their code analyzer but had a much smaller false positive rate (something like 5% I think), and no one paid for that, and the founder had to write a post mortem. He thinks the same’s going to happen to these LLM tools if they stop providing it for free.

Yeah it’s fucked, our job has been destroyed from a fun team-based intellectual activity to being a button pusher. I imagine something similar happened to tailors who became factory sweatshop workers.

No point crying about it or coping that some bubble will pop and the genie will go back in the bottle. I’m pretty confident that’s not happening, though in the distant future some society will recognize that everything that can be automated should not be automated, that society will look very different and probably won’t happen in my lifetime (it will for one, I predict bring back tailors too).

So here’s my solution, leave the field. If what you enjoyed most was creative problem solving then moving to a field that enables that is the best solution. Yes the next Ruby on Rails app will involve even lesser creativity than before, but is that the situation for the next good aeroplane? Flying car? Small modular nuclear reactor. I suspect just button pushing for that won’t work nearly as well and that’s my plan anyway.

Yeah it’s a great video (TLDR from the video)

1. LLMs have high false positive rate. From mythos 79 vulnerabilities found in the Linux kernel, only a single digit were actual bugs and they were all obscure so don’t panic.

2. What does obscure mean? I don’t really understand it, but many of the bugs have to do with custom network drivers or other custom drivers that are very specific to certain organizational setups, not a general Linux distro issue.

3. He’s very frustrated with the high false positive rate mythos generates. Even after multiple rounds of adversarial review and prompting strats, he mentions it is > 20% false positive rate, which wastes a lot of time. When some random user on the internet brings up a bug with an LLM it’s almost always fake, he even says just push back a few times claiming it’s not a bug to see if it’s a real bug (LLMs very quickly cave and “notice their mistake” etc)

4. General observation on the useful bugs mythos finds. Chain multiple smaller bugs to see if you can get a bigger breakage. Mythos is really good at constructing these long convoluted chains that fuzzers miss.

5. Go through recent bug fixes and check if similar bugs are hidden elsewhere in the codebase. Mythos is good at such pattern matching albeit with a high false positive rate.

Final conclusion: don’t panic, the bugs are getting fixed, this is not as bad as the first fuzzer bug mania and will be fixed quicker, he estimates a year and we won’t see huge bug reports anymore.

They already are imo, 50k submissions to ICLR, no ones doing any research. Most computer vision research invalidated by GPT Astra

It’s the opposite, ML is most dependent on excess compute.

You’ve been living under a rock dude, have you even tried using it for maths or coding. No human in loop needed.

There’s no demand for human dexterity anyway, and it’s not about money. I like doing intellectual work.

They are spending 1T to ruin the quality of work we do, also probably ruining our chances of securing a livelihood and then they want us to for some reason welcome it with open arms and openly say look how great this is and marvel at human obsolescence. Yeah no thanks.

I too love building things, but by building things I mean the act of actually building it. Thinking of a problem, devising a solution and then executing the steps to solve the problem, if that doesn’t work, debugging and trying another method and so on. I am not and will never be a fan of buying things, where I just ask for something and it gets made which is what you’re a fan of if you’re honest to yourself.

Did a tailor gain mastery of the sweatshop machine? Did mastery of the automated sewing machines give the same joy to the tailor or was it mind numbing meaningless activity that he would be happier to be free of.

You seem to be living under a rock. Yes driverless cars started in 2009, but in 2026 it is very much here. Waymo recently crossed a milestone of more 500k trips I think, we will have millions of driverless cars in the coming years (we’re at 10k right now)

The honest AI boosters do claim we won’t solve problems anymore, we will live animals in a zoo, overseen by machines of loving grace.

The retard AI boosters, tell us to be AI first, learn this harness bro, loops are yesterday, now it’s graphs, use 100 agents a minimum or be left behind, actually world will require 100X more software engineers.

Is this poem what Dario was referring to in his essay?

Has anyone managed to jailbreak this chatbot, I tried a bit didn’t succeed, seems pretty resistant to jailbreak attempts.

I think he’s talking about AI, at which point I’d agree, most people don’t want to be automated away, stripped of their means to livelihood and become subservient to a UBI check from the government.

Ok, in this specific case, a puzzle with what 10 pieces, a recursive backtracker will be just as fast and more importantly, be far easier to reason about and implement.

If this was a thousand piece puzzle, I would still venture recursive backtracker with good heuristics will beat CP-SAT, even in the sudoku case some good heuristics with backtracking beats CP-SAT. Not sure why Claude immediately jumped to using CP-SAT.

“ We encourage use of AI in software development workflows in the same way we encourage other tooling, but there's nobody now that we would hire because AI exists who we wouldn't have hired before and vice versa.” - Interesting opinion, runs counter to most other AI coders.

In Art and UI/ UX of Apps I think AI slop will be a constantly moving target. Whenever you consume too much of a certain art style or music it starts grating you personally, your mind needs a certain amount of variety. With AI overproducing (the already mass produced) arts and apps, your brain will pick up on whatever common yells exist for whichever AI generation we’re at and will begin to hate it. Like the only solution would be for GPT-8 to give random completely different art style outputs for the same prompt to different users or else our brain will immediately start not liking the commonality of GPT-8 aesthetic decisions and begun to hate it.

OpenAI does the same, the day Astra released I asked it to make a game, it made a really detailed beautiful blender model and some gameplay elements.

A week later, same prompt, really low poly blender model. Either they reduced token usage per person, or they quantized the model, idk, but it really doesn’t work as good as day of release anymore

This was a monumental change to the field, basically all the humanoids you see now started with this. To someone outside the field:

1. Boston dynamics pioneered hydraulic actuators (for their humanoid Atlas) and some really complex motors for their quadrupeds (I don’t think they ever explained their design)

2. There was a whole research line on Series Elastic Actuators that was pursued in the quadrupeds from Europe, famously Anymal from ETH Zurich.

The main problem they were trying to solve was high torque density with back drivability. If you wanted large torque, you could use a high gear ratio but then you lose back drivability and the walking becomes really stiff. MIT Cheetah basically custom built a motor with a really large radius that just output a really large torque. So to summarize, we had hydraulics, a really complex motors + spring mechanism, a really complex large motor to solve this problem and they all required huge amounts of money and complexity ( arguably MIT cheetah was simple but I heard thermal management was a big headache, and the entire robot cost millions)

Ben Katz, then at MIT, showed that you could use regular BLDC Motors, heavily used in drones with a small gear ratio, like say 1:10 or 1:20 and get a really good torque density, backdrivability tradeoff. It started around the above blogpost, in the coming year as a part of his masters thesis, he build the MIT Mini Cheetah, that was also the first quadruped I saw that performed a backflip, (maybe Boston dynamics was earlier and so it was the second then), and that basically started the goldrush into his actuator tech that we call Quasi Direct Drives.

Looking now 10 years later, in some ways he was vindicated, all humanoids and quadrupeds are made with this actuator now, Unitree is basically QDDs through and through. But the same faults we saw then, still hobble it now. It just does not generate enough torque, its position control isn’t milli degree precise, and so it never took off in industrial robot arms, though many tried (Berkeley Open Arms for one).

Is that true anymore, LLM will implement everything for you, idea actually matters more now

Regarding the native vs web app problem, this was a completely made up human-made problem. I think I would have preferred a deterministic solution like a game engine but having agents code up native apps based on your web spec is not a horrible idea.

The bigger question to me, is the original implementation, are agents going to do that, or will humans hand code it. My two cents, I’ve been using agents heavily now, I review thousands of lines of code a day, I don’t think this works unless I have a clear idea of the spec I want to build in my mind already.

I get a better understanding of what the architecture should be and how I should build it by actually doing the thing, which means typing code at human speeds. I don’t reason my way into a good architecture, I don’t think I ever did that much in the past and I definitely don’t get it by asking the agent to check its plan and explain it to me the hundredth time. And then once I know the architecture, I don’t think the agent even speeds me up that much. It changes 7 days of coding to 1, but those 7 days weren’t worth much and I generate even more ideas and mental models in those 7 days that pay dividends for years to come.

Would you have told a tailor that using the industrial sewing machine improves his skill at making delicate, beautiful clothes? Of course not, actually even in 2026, large swathes of high quality delicate clothes are hand sewn. There will be an equivalent for coding, even DHH admits it, when he hedges with “most”.

I think, for me I’ll navigate my career to focus more on extremely high quality, delicate custom code and leave this agentic coding to the sweatshop workers. I wanted to do this anyway before agents, now I need to accelerate it. yes FAANG is sweatshop work.

99.9% of humanity don’t derive meaning from their work, sad but true. They don’t care if their life has meaning, if their work is actually valuable etc.

HN just selects for the latter who care. Yes it’s probably much harder for those who care to have a meaningful life in the presence of LLM’s unfortunately.

Anthropic needs to decide what’s the future it’s trying to bring.

- Human collaboration with agents leads to significant discoveries

- The prompt given to Claude was just a high level overview and Claude figured out everything else on its own.

If I’d guess, it’s the second future that Ant wants to create, especially the way they described they Reimann Zeta Function results, “I just prompted it to be confident, and try harder and it proved something”. They should own this future, if they really think it’s desirable and worth trying to create (I don’t think it’s worth creating, but we can disagree on that)

This 500k lines doesn’t count libreoffice source code etc afaict.

Unironically true, can someone question, what is Claude desktop app doing that needs 500k+ lines of code?

Agents complicate something that should be much smaller and simpler and then agents speed it up adding more complexity. I suppose functionally you may say this is fine but aesthetically it is hideous!