This is about to change very soon. Unlike many other domains (such as greenfield scientific discovery), most coding problems for which we can write tests and benchmarks are "verifiable domains".
This means an LLM can autogenerated millions of code problem prompts, attempt millions of solutions (both working and non-working), and from the working solutions, penalize answers that have poor performance. The resulting synthetic dataset can then be used as a finetuning dataset.
There are now reinforcement finetuning techniques that have not been incorporated into the existing slate of LLMs that will enable finetuning them for both plausibility AND performance with a lot of gray area (like readability, conciseness, etc) in between.
What we are observing now is just the tip of a very large iceberg.
These humanoid robots making cars is always a bit disappointing. If you look at videos of an assembly line the humans do manually dexterous complex tasks like bolting in car seats which are move in by hand, or helping place windshields into position.
The problem humanoid robots solve would be addressed better by simply altering the design of the car to not require humans to do that stuff.
What we actually see is humanoid robots deployed to do tasks that can already be done with simpler robot arms...
That's not how a smart business runs and that's now how Nvidia operates.
Jensen is smart. He's gone through over 30 years of tech cycles.
Nvidia actively commoditizes the LLM models. Look at Nemotron. They've avoided making a SOTA model solely to keep the hyperscalers (aka crack addicts) coming back for more GPUs.
As soon as the bubble bursts, they can release some open weight NemoMambaDiffusiontron and keep folks buying GPUs to run the damn thing.
It still wouldn't be smart to do so, as this would fall into the common business pitfall of thinking you could easily do the next stack layer of work.
The edu discount is often $50 or $0 on the lower cost products. E.g. that new M4 iPad Air from earlier this week is also $599 but has a $50 discount on the us-edu store. It's pretty sweet they kept the neo at $100 discount even though edu is one of the primary markets for this.
The question is why was it so urgent to bring this up right now. The answer is that the US Navy has carrier battle groups stationed off the coast of Iran right now.
US F22 fighter jets have been spotted on Israeli airfields by Chinese commercial satellites in the past 3 days.
A full scale strike on Iran is imminent and they want to leverage LLMs for data analysis and quite possibly automated target selection.
Literally the SkyNet scenario both researchers and frickin pop culture have warned about for years.
I thought it sounded very conspiratorial (can you even integrate a new model that quickly?) until I learned that Israel launched strikes on Iran literally as I was typing out the comment. I bow in respect to your analytic skills.
This is really about the imminent strike on Iran which is now super telegraphed. They are gonna use ChatGPT for target selection, and the likely outcome is that it will fuck things up and a bunch of civilians are going to die because of this decision.
When this happens, Altman will go from being merely a drifter to having blood on his hands.
The Iranian Revolutionary Guard Corps has over 200,000 personnel. There are thousands of putative SAM sites and MANPAD launch sites. The amount of data to crunch is significant.
Alternatively, the DoW is simply incompetent and Trump or Hegseth wants to use AI to draft war plans.
The Iran situation is unique. If it is true that Epstein was part of a blackmailing operation run by Israeli intelligence, then the time to act is limited. It may only be a matter of time before the US-Israel special relationship begins to deteriorate, especially as the House of Representatives starts digging into what was going on.
For hardline right wing Israeli government officials who would be privy to such information, the window of time to leverage to US to enact regime change on the Islamic Republic is closing. The survival of Israel over the long run really depends on not having a hardline Islamic regime in Iran developing nuclear weapons. Things like AI safety and US elections are secondary to such prerogatives. The question for voters in the US is whether it really is worth it to the average US citizen to shed blood and tax dollars for this stuff.
I hope there can be a peaceful regime change in Iran and that there will be peaceful relations with Iran and Israel in the future. But damn I wish things could go back to normal with our US political system once this is all settled.
Vibe coding isn't so much a movement as a big fat tool that was air dropped from space after the megacorps decided to dump billions of dollars into LLMs and LLM companies.
It's like comparing Christianity to water wheels or gay pride to to the Saturn V rocket. It's just not really analogous in any way.
I do agree with the author about commoditization, however.
The most likely outcome is that software will be commoditized and software developers commoditized even harder. If we still need software engineers to prompt, you'll find plenty of people in India able to do those tasks, not necessarily with great quality until they too are replaced by better AI.
This whole situation inspired me to actually dive harder into Maker type stuff such as learning how to design PCBs, but one thing I found is that this TOO is very close to being automated by AI. To actually get hardware made, even prototyping PCBs, you NEED to go to China, and the Trump tariffs cut into the cost of doing these activities hard.
The outsource-prompting-to-India stage will almost entirely be skipped (it already has been).
Developing nations that were looking to tech to climb the economic ladder, are watching that ladder be pulled up.
Most of the upside will go to the US and China. Europe is lagging shockingly on AI spend, they're extremely far behind (but with constant plan announcements). If you didn't know any better, you'd think Europe believed the year was 2010.
Circuit design has been on the cusp of being automated ever since there were computers. It's been over twenty years of autolayout tools and they're still not very good.
Maybe you could research how to make your own PCBs? It can be done at home with a little equipment and then you can offer it as a service to others.
The thing is, the cost of making them at home is both laborious, poorly reproducible, and lower quality than getting it done at JLPCB. In the old days, it was done with an exacto and rubylith. Today it requires a laser setup for about $4000 and you still need to manually apply and remove a resin layer. It's mostly not worth it to do things this way, and you lose the ability to have more than 2 layers in your board.
The bigger issue with PCBs is that even with a nice prototype, actual manufacturing needs to be done in Shenzen for any sort of cost competitiveness if you want to step outside of the hobby realm (just as the author of this essay stated).
It's really, really hard to see where the USA stands in the value chain any more once LLMs have been deployed. If all the physical manufacturing lives in China, where the manufacturing supply chain also lives, and Chinese AI companies can very very easily distill US LLM models, the two remaining US advantages is actually just the dollar's role as reserve currency --- something that crypto bros and the US president is working on eroding at a record pace, and the fact that this overvalued dollar gets all the smart people in China and India to emigrate to the US -- also becoming politically less viable.
This means an LLM can autogenerated millions of code problem prompts, attempt millions of solutions (both working and non-working), and from the working solutions, penalize answers that have poor performance. The resulting synthetic dataset can then be used as a finetuning dataset.
There are now reinforcement finetuning techniques that have not been incorporated into the existing slate of LLMs that will enable finetuning them for both plausibility AND performance with a lot of gray area (like readability, conciseness, etc) in between.
What we are observing now is just the tip of a very large iceberg.
reply