lexandstuff

Prompt engineering isn't really that important anymore imo. If you're using a reasoning model, you can see if it understood your request by reading the reasoning trace.

I've never had a job like that. My job has always involved helping my company, not just figure out how to build something, but what to build. We typically collaborate on a few ideas and then go away, let them percolate in our brains, before coming back with some new ideas to try. The whole point of the Agile Manifesto is that we don't know what to build in the first place.

Sometimes my boss has asked me to do something that in the long run will cost the company dearly. Luckily for him, I am happy to push back, because I can understand what we're trying to achieve and help figure the best option for the company based on my experience, intuition and the data I have available.

There's so much more to working with a team than: "Here is a very specific task, please execute it exactly as the spec says". We want ideas, we want opinions, we want bursts of creative inspiration, we want pushback, we want people to share their experiences, their intuition, the vibe they get, etc.

We don't want AI agents that do exactly what we say; we want teams of people with different skill sets who understand the problem and can interpret task through the lens of their skill set and experience, because a single person doesn't have all the answers.

I think your ex-boss Mike will very soon find himself trapped in local minima of innovation, with only his own understanding of the world, and a sycophantic yes-man AI employee that will always do exactly as he says. The fact that AI mostly doesn't work is only part of the problem.

Re the last sentence, is the answer that more people will die in aviation disasters?

People seem to have forgotten how good the 2023 GPT-4 really was at coding tasks.

Except that, all Deepseek models so far have been trained on Nvidia hardware. For Deepseek v3, they literally mention that they used 2,048 NVIDIA H800 GPUs right in the abstract: https://arxiv.org/html/2505.09343v1

"We added a placeholder since there are rumours swirly, and they're from reputable sources - coincidentally the timelines for releases (around 2 months) align, and it's on a Monday, so it's highly likely.

But it's all speculation atm!"

https://www.reddit.com/r/LocalLLaMA/comments/1kvoobg/comment...

Presumably, for this concept to stick with you your whole life, you'd have to have heard of "Borges"? From Googling that name, it appears the author you're referring to died in 86. Why would anyone expect him to win the current year's Nobel Prize in Literature?

That is amazing. But I had a similar experience when I first taught my mum how to Google for computer problems. She called me up with delight to tell me how she fixed the printer problem herself, thanks to a Google search. In a way, LLMs are a refinement on search technology we already had.

One of the realisations I've had recently is that the AI hype feels like another level from what's come before because AI itself is creating the "hype" content fed to me (and my bosses and colleagues) all over social media.

The FOMO tech people are having with AI is out of control - everyone assumes that everyone else is having way more success with it than they are.

Funny, I would give the absolute opposite advice. In my experience, the use of agents (mainly Cursor) is a sure-fire way to have a really painful experience with LLM-assisted coding. I much prefer to use AI as a pair programmer, that I talk to and sometimes get to write entire files, but I'm always the one doing the driving, and mostly the one writing the code.

If you aren't building up mental models of the problem as you go, you end up in a situation where the LLM gets stuck at the edges of its capability, and you have no idea how even to help it overcome the hurdle. Then you spend hours backtracking through what it's done building up the mental model you need, before you can move on. The process is slower and more frustrating than not using AI in the first place.

I guess the reality is, your luck with AI-assisted coding really comes down to the problem you're working on, and how much of it is prior art the LLM has seen in training.

That's similar to asking why does everyone need a GP? Most people experience some kind of mental health challenge in their life.

Your 2nd question is much more interesting to me. Why is it so hard to find a good therapist?

It's no surprise to me that people are turning to ChatGPT for therapy. It does a decent enough job and it doesn't have a 2-year waiting list, or cost $300 a session.

Because the original point of it was to distinguish journalists and public figures, so you could tell imposters apart from real people. Now its purpose is to show who has a premium subscription. Totally different feature using the same name.

Related note: Karpathy is really good at naming things. Seems whatever name he comes up with for new things sticks.

He was originally testing the Automatic Emergency Braking feature, but it couldn't stop in time for a single obstacle, so he cut them some slack and used Autopilot instead.

Also up 0% from 4 years ago.

What "potential peace deal" are you talking about? Russia has no interest in peace while they have so many allies in the Whitehouse. Meanwhile they continue to bomb and kill Ukrainians.

This is 100% my experience. I appreciate a manager who can jump into the codebase to fix the small stuff: typos, lint issues, updating minor dependencies, etc., unblocking devs from doing the main work. I like when they have some sense of the reality of the codebase, as you put it, and know who is actually contributing vs bullshitting.

The worst managers I've ever had were the so-called "technical" managers who had never looked at the code. They were often involved in technical decisions, but their opinions were entirely based on vibes. Since they were a manager, people felt obliged to listen to their input, even if it was disconnected from reality.

Either: a) be completely non-technical, and make sure you have a technical leader on the team who you trust, who does know the code or b) get involved in the code, enough to support and unblock your team.

These high-quality music models require pirating many, many terabytes of music. Torrents are the main way to do it, but they likely scraped sites like Bandcamp, Soundcloud and YouTube.

AI music is a weird business model. They hope that there's enough money peddling music slop after paying off the labels (and maybe eventually the independent music platforms) whose music you stole. Meanwhile, not even Spotify can figure out how to be reliably profitable, serving music people want to hear.

Also, if you're not from the USA, you're more likely to associate it with the beloved Rowan Atkinson sitcom.

To the absolute contrary. They do it because of the contempt they have for the average US citizen. These people now in power believe their interests come first because, in their minds, they are inherently superior to most people by birthright.

Another example of DeekSeek copying straight from OpenAI's playbook [1] [2]

[1] https://www.reuters.com/technology/cybersecurity/openais-int...

[2] https://openai.com/index/march-20-chatgpt-outage/

We've been seeing success using it for LLM-as-a-judge tasks.

We set up an evaluation criteria and used o1 to evaluate the quality of the prod model, where the outputs are subjective, like creative writing or explaining code.

It's also useful for developing really good few-shot examples. We'll get o1 to generate multiple examples in different styles, then we'll have humans go through and pick the ones they like best, which we use as few-shot examples for the cheaper, faster prod model.

Finally, for some study I'm doing, I'll use it to grade my assignments before I hand them in. If I get a 7/10 from o1, I'll ask it to suggest the minimal changes I could make to take it to 10/10. Then, I'll make the changes and get it to regrade the paper.

Sounds like your CTO is dangerously incompetent and isn't actively using the AI tooling they are endorsing.

I don't know what the future holds, but developing software in the present is amazing. I've always loved writing software, but I love writing software with AI even more.

All the boring stuff, like converting data from one format to another, is a prompt away. All the annoying scripts you know how to write, but cbf, like conditionally editing files across directories, is a prompt away.

Plus, I feel secure with my AI buddy. It's like having a helpful lead who's always up to help me debug, give me a second opinion, or show me what I'm missing. I feel like I can work in any language I want, and learning new things is so much easier with an AI tutor.

The difference is that you didn't need to scrape millions of videos from YouTube with residential proxy network scrapers to train yourself.

Good article. A very relevant paper worth reading is Programming as Theory Building: https://pages.cs.wisc.edu/~remzi/Naur.pdf.

Programming is not just about producing a program, it's about developing a mental model of the problem domain and how all the components interact. You don't get that when Claude is writing all your code, so unless the LLM is flawless (which it likely never be on novel problems), you won't understand the problem enough to know how to fix things when they go wrong.

My guess is that it isn't available because the training data they stole occasionally leaks into the outputs.

The NotebookLM podcast feature is so unbelievably good! I've been feeding it random papers I've always wanted to read, and it's giving me these entertaining high-level summaries. The banter is almost addictive. Also, reading the papers in detail is quite easy after you understand the topics.

The trade-off was that you had to burden your friends to take the photos, spend your own time collecting the images from them after your wedding, research AI services, and oversee how AI touches up your photos.

You traded money for your and others time. Nothing wrong with that, but AI hasn't changed anything except the crowdsourced images look a little better, maybe?

I would rather pay to get the 30 great photos ready to go - save me the hassle. Plus, a photographer will coordinate the key wedding photos with the various family and friend groups and will help find the ideal scene for shots.

Of all jobs, wedding photographers are relatively safe from generative AI.

Yes. The dead giveaway is multiple paragraphs of bullshit where one sentence would suffice. People ain't got time for that.