If you want to generate a heatmap of existing text, you will have to take a different approach here.
The naive solution I could come up with would be really expensive with openai, but if you have an open source model, you can write up custom inference that goes one-token-at-a-time through the text, and on each token you look up the difference in logprobs between the token that the LLM predicted vs what was actually there, and use that to color the token.
The downside I imagine to this approach is it would probably tend to highlight the beginning of bad code, and not the entire block - because once you commit to a mistake, the model will generally roll with it - ie, a 'hallucination' - so logprobs of tokens after the bug happened might only be slightly higher than normal.
Another option might be to use a diffusion based model, adding some noise to the input and having it iterate a few times through, then measuring the parts of the text that changed the most. I have only a light theory understanding of these models though, so I'm not sure how well that would work
I've spent the last few months using Claude Code and Cursor - experimenting with both. For simple tasks, both are pretty good (like identifying a bug given console output) - but when it comes to making a big change, like adding a brand new feature to existing code that requires changes to lots of files, writing tests, etc - it often will make at least a few mistakes I catch on review, and then prompting the model to fix those mistakes often causes it to fix things in strange ways.
A few days ago, I had a bug I just couldn't figure out. I prompted Claude to diagnose and fix the issue - but after 5 minutes or so of it trying out different ideas, rerunning the test, and getting stuck just like I did - it just turned off the test and called it complete. If I wasn't watching what it was doing, I could have missed that it did that and deployed bad code.
The last week or so, I've totally switched from relying on prompting to just writing the code myself and using tab complete to autocomplete like 80% of it. It is slower, but I have more control and honestly, it's much more enjoyable of an experience.
I do like the control you have over organizing. The only thing I was shocked was I couldn't revert a series of changes (checkpointing) like you can can in Cursor. Sometimes the LLM does it a different way I dont like, so I want to go back and edit the prompt and try again.
I grew up in a town called Seymour, which is in the same county as Gatlinburg/Pigeon Forge. It's Dolly country. She's a hero - she's done so much for her community. Much of the money you spend at Dollywood goes to help so many public causes. One of them is the Imagination Library, which gives out free books every month from birth to 5 years of age. I was one of the first families that got to take a part in it. I certainly don't remember much from when I was 5, but I do remember getting those books. I imagine it had a positive impact on my growth.
We'd go to Dollywood a few times a year - she would give out free tickets to people who worked in Gatlinburg to go. It's really well run, and their water park is great too. Growing up, we'd ride the train when we visited. While I didn't appreciate it much as a kid, when I grew up I realized how awesome of an opportunity that was.. I moved away from Tennessee about 12 years ago, one of the biggest things I have missed is Dollywood and their big steam train.
I can't really show an interactive demo, but my team at my day job has been fine tuning OpenAI models since GPT-3.5 and fine tuning can drastically improves output quality & prompt adherence. Heck, we found you can reduce your prompt to very simple instructions, and encode the style guidelines via your fine tuning examples.
This really only works though if:
1) The task is limited to a relatively small domain (relatively small could probably be misnomer, as most LLMs are trying to solve every-problem-all-at-once. As long as you are having it specialize in a specific field even, FT can help you achieve superior results.)
2) You have high quality examples (you don't need a lot, maybe 200 at most) Quality is often better than quantity here.
Often, distillation is all you need. Eg, do some prompt engineering on a high quality model (GPT-4.1, Gemini-Pro, Claude, etc.) - generate a few hundred examples, optionally (ideally) check for correctness via evaluations, and then fine tune a smaller, cheaper model. The new fine tuned model will not perform as well at generalist tasks as before, but it will be much more accurate at your specific domain, which is what most businesses care about.
I'm in my mid-thirties and most of my friends have ditched Facebook. I didn't really realize this until when I used it to create an event for a house party... I was somewhat surprised that only 2 people out of 15 even saw it. I ended up resorting to good old text message and that worked, but it was tedious. Not sure how popular this will become, but having a social-media-less event invite/broadcasting system would be nice, and having one that most people with an iPhone have access to covers much of my friend base
The story is that these young engineers built a 500$ drone that consumes this kind of data to do mapping. In 24 hours for a hackathon no less.
If the US government didn't already have this kind of tech, they would spend millions just for the same prototype they built. And probably tens or hundreds of millions for a final product.
Is 2.9% and 6.8% acceptable or even 'good' levels, even for the tech industry? I'm not quite sure. A 6.8% performance-related quit/termination seems pretty high from my limited anecdotal experience.
Edit: now that I'm thinking about it, I don't think I worked at a tech company where people were fired/asked to resign more than 2.5x than those that resigned on their own.
I don't know about LinkedIn, but a comparable company would be Indeed - something like 12-15k employees. The majority of the company is sales & marketing. Their customer base is basically every company in the world, so having each sales person nurture a few dozen to a few hundred relationships is not uncommon.
If you sign up today and host a job on Indeed, you will only be charged for applicants you approve. You have a day or two to reject candidates. We charge more money per candidate that you approve, instead of charging you per apply or click.
Internally, Indeed watches a ton of metrics, but our primary metric is how many people we ultimately helped get a job. We rarely release a feature unless we can prove it, with hard data, that it does. Secondary metrics monitor things like jobseeker and employer efficiency (how easy is it to find a mutual match.) In the last two years, the pricing model, the recommendation engine, everything is completely different.
I'm sorry, but I get irked whenever I see a CEO saying they take full responsibility but then accept no consequences. There is no impact on their compensation, their equity, or their life other than the "emotional scar" left behind of having to let people go. They will absolutely expect the remaining staff to pick up the game from the loss of such a big chunk of the company. It kind of feels like the letter is more or less "we know you are hurt, but so am I and think about my feelings too"
I know I'm being pessimistic, but the CEO isn't the one having to spend the next few months trying to find a good job when everyone is on a hiring freeze.
As an interviewer, this is so incredibly frustrating - we don't change our questions often and because of that the questions and answers are all over these forums. With that said, it is incredibly easy to spot someone cheating - they often write the most perfect optimal solution from start to finish, helper functions first, often with the same function names as the forums themselves. The trick I've learned is to ask "why" - "why did you decide to use a linked list for this problem?" - the real cheaters often freeze up and can't give an explanation at all. It's rare, but still too common.
If you've seen the question/answer before just say so! I will totally appreciate the honesty and it goes a long way.
The naive solution I could come up with would be really expensive with openai, but if you have an open source model, you can write up custom inference that goes one-token-at-a-time through the text, and on each token you look up the difference in logprobs between the token that the LLM predicted vs what was actually there, and use that to color the token.
The downside I imagine to this approach is it would probably tend to highlight the beginning of bad code, and not the entire block - because once you commit to a mistake, the model will generally roll with it - ie, a 'hallucination' - so logprobs of tokens after the bug happened might only be slightly higher than normal.
Another option might be to use a diffusion based model, adding some noise to the input and having it iterate a few times through, then measuring the parts of the text that changed the most. I have only a light theory understanding of these models though, so I'm not sure how well that would work