[X6S1x6Okd1st]

He started learning lean4 with the help of GPT4 just at the start of the month:

https://mathstodon.xyz/@tao/111208692505811257

Many of his mastodon posts this month have been about his learning progress.

Certainly an interesting case study of how LLMs can accelerate the work of even of the most extremely successful people

[mensetmanusman]

I have found that good communicators that don’t code can quickly make functional automation.

Interestingly, LLMs may end up contributing to more inequality if only the highly skilled can leverage them effectively.

[anonylizard]

LLMs CURRENTLY favor the more curious and open individuals (High on openness in big 5 scale). Half of the population is not open, does not want to try new things, unless it leads to very direct and proven benefit.

However, over time, the overwhelming benefits of using LLMs will be well understood, and these ladder climbers will absolutely master LLMs, no matter their intelligence. People can become experts at taking exams despite how boring and soul sucking that can be, let alone using something way funner and useful like LLMs.

[marshray]

My friend had never written anything more than an Excel formula a few months ago and now he's using GPT-4 to write very nontrivial Python applications and automate large parts of his job.

I (having 30 years experience as a professional Software Developer^TM) am begging him to teach me his techniques.

Now that you mention it, I met him and we became friends in large part due to his communications abilities.

[an_aparallel]

5 months ago - a friend wrote me a python script and sent it to me...i couldnt get it to work. Used phind.com to explain what to do...it worked out my windows environment variables needed to be changed, told me how to structure a folder schema to place the src script...mindblowing stuff. And i have been using it - when it turn the same friend told me to write a similar script in python myself...it has been amazing to forego stackoverflow/google to get answers. When the GPT-4 model kicks in - my questions are answered with beautiful clarity...whether or not this technology "hallucinates" or not...without a doubt it's empowered me to learn to program. If it's helped me out of 50% of my ruts - i count that as overwhelmingly positive!

[heavenlyblue]

I am not going to visit the website you just posted here unless you explain how the mechanism works, otherwise I will assume it's just crappy advertisement in this post.

[hobs]

Phind is just prompt engineering, as is anything sitting on top of anything OpenAI is building right now - if you are doing Q&A then maaaaybe it does some resource augmented generation as well (say, from Stack Overflow or other technical resources)

Phind is often recommended to me by programmers who find it produces "better" results than a naive GPT-4 session, but I don't know that anyone has done any real world testing.

[an_aparallel]

as explained before me phind is a website built ontop of GPT-4 + a free tier (which is what i use) - which uses what they call the Phind Model...though it changes often.

I understand your skepticism (because the internet is fully broken right now/and forum posts are more or less paid for content at this point)

but...its a bit funny you'd like someone who just explained they're a beginning programmer who got into programming because an LLM helped remove insane amounts of friction from the learning process to "explain a mechanism" to them.

[astrange]

Nobody knows how LLMs work, much less LLMs behind an OpenAI server API, so that's a hard question.

[melagonster]

sorry, everyone is losing their job. we have no chances of living in future.

[strikelaserclaw]

gpt4 is amazing, i rarely use google as as starting point for my programming related queries these days.

[kaba0]

Google has actively become catastrophically bad to the point that it will ignore the only special keyterm I deliberately looking for for more general dumb results.. like it is literally unusable - I used to be able to find that very specific stackoverflow answer I have read years ago, or some barely read blogpost, and now I’m happy if it actually finds the website if I search for the domain name..

GPT would be good as a search engine, but who would want a search engine that stopped indexing a few years back? Also, it is not good at ultra niche topics, which would be the whole point of a search engine.

[justinclift]

Yeah. DuckDuckGo also seems similarly bad these days too. :(

[RealityVoid]

Huh, I can't seem to get in the groove of using it, maybe I'm old or something, but it annoys me all the subtle ways it's wrong and I feel I have a much better grasp if I think through it myself supported by Google.

[popularonion]

I get a lot of mileage out of ChatGPT just treating it like an intern who turns around work instantly. You don't expect interns to write perfect code, but they can save you a ton of time if you set them loose on the right problems.

For any relatively simple task I can say "Write a Python script to do X" and it will almost always spit out working code, even if it has subtle mistakes. Fixing mistakes is fine and part of the process. I don't have to read StackOverflow posts saying "Do you really want to do X?", or sift through documentation that follows the author's approach of how they want to introduce the material but doesn't directly address my question.

[amelius]

Not OP, but I always want to read the code generated by chatgpt before I run it. And I dislike reading other people's code much more than writing it myself.

[danenania]

Especially now that the training cutoff is being moved up to April 2023. Questions requiring more recent results were the main ones I’ve been going back to google for.

[shusaku]

Part of this is that ad infested AI generated blog spam is flooding Google! But it’s also my go to. I also really liked GPT to bring me up to speed on a libraries I’ve never used.

[colinhb]

Agree in part but also I think Terry is such an outlier (though also generous and humble) that it’s hard to extrapolate from this example to a more general case

[lulznews]

They’re an easy 100x for the elite. Top engineers are now 10000x’ers.

[voxl]

The only benefit Terrence Tao has received is in correcting syntax. He has mentioned that he has gotten (in my opinion better) benefits from talking to existing Lean4 specialists.

[rowanG077]

I don't think you realize what you are saying here. I agree that it is a large boost but 100x is just too ridiculous to take serious. Do you really believe an engineer can now finish in 20 hours what would have taking them a year before?

[ghshephard]

As someone who uses ChatGPT4 pretty much nonstop all day (I have a monitor dedicated to it) - it's almost certainly a 10-20x in fields I have no knowledge of (writing in languages I've never used, dealing with OS/Kernel/Network constructs that are unfamiliar to me) - I can knock off in an hour what might have taken me a day or two previously - but I don't think I've ever had a task where I could say that I completed in 1 hour what would have taken 100 hours - though I would love to hear of counterclaims from people who have been able to do so - definitely not saying impossible, just that I haven't had that experience yet.

[pradn]

The idea of the 10x programmer can mean 1) someone who produces a ton of code quickly 2) someone who can solve seemingly-intractable problems 3) someone who’s presence on a team improves everyone’s productivity quite a bit 4) someone who chooses technical decisions that save a ton of time down the line.

[amelius]

(1) is the generally accepted meaning. The others need different words.

[hobs]

Disagree, those are all valid interpretations, and depending on your experience you will have a vastly different understanding.

To me, the first one is the most basic and frankly, most silly definition of a 10x programmer - produces more code? really? Code sucks. Nobody needs more code, people need solutions. Solving a problem with no more code, or understanding it so a _tiny_ change solves all problems? Way better than adding code, actually much harder to synthesize and produce new results. Way more efficient.

Now can you remove things and solve the problem? Realize you can build an even more simple and generic system that solves your problem? Even more amazing.

[firedaemon]

I have always used it to mean “solves problems 10x faster” not “produces 10x code”.

The latter would be a bad thing.

[kaba0]

I would argue that it only helps in very beginner questions as it is beyond useless on anything more complex.

Like if the answer is not a leetcode/CRUD code sample you can easily find on the internet either way, it can’t do anything useful. Like, my problems mostly constitute things like reasoning about the correctness of a given change in the context of the whole codebase, its threading model, etc. It will at best regurgitate some bullshit like “race conditions can be this and that” here, not helping at all.