arkmm

This is a really great explanation.

IMO if you think you can sell to users within the niche, you can publish a blog post of benchmarks and that'll serve as strong technical marketing for your niche.

It also keeps open the option to sell to an incumbent (possibly helps maximize the value of that option as well).

Missing the California high-speed rail on their list of examples.

When do you think fine tuning is worth it over prompt engineering a base model?

I imagine with the finetunes you have to worry about self-hosting, model utilization, and then also retraining the model as new base models come out. I'm curious under what circumstances you've found that the benefits outweigh the downsides.

Automating applying to jobs makes sense to me, but what sorts of things were you hoping to use Operator on Amazon for?

A bit of a minor detail, but this piqued my interest "DOM-aware browser integration" - could you say a little more?

Not affiliated, but someone's already working on that for you: https://www.realavatar.ai/

This is cool! Also have found the Playwright MCP implementation to be overkill and think of it more as a reference to an opinionated subset of the Playwright API.

LinkedIn has this reputation of being notorious about making it hard to build automations on top of, did you run into any roadblocks when building your personal LinkedIn agent?

This is an interesting take, thanks for sharing. I wonder how someone should adjust their blogging if they believe their primary audience will be LLMs.

What are some examples where you've provided the LLM enough context that it ought to figure out the problem but it's still failing?

Did you try recalibrating your cameras? After a windshield replacement my HW3 FSD camera calibration was off and it was similarly terrible like you were describing (going way too close to curbs), but then I recalibrated on a straight highway stretch and since then it's been amazing. Over the past few thousand miles, I only take over for parking and getting out of parking lots at this point. (Also the occasional intervention if I want to take a different route that I think might be faster.)

Congrats! You never forget your first.

Very cool to see an early stage company doing this! I always hear that healthcare has a lot of red tape to handle so it's hard for startups to operate without tons of VC funding.

How'd you guys find your initial users and figure out rehab clinics were a good place to start?

I think most code these days is boilerplate, though the composition of boilerplate snippets can become something unique and differentiated.

Neat - didn't realize there were affordances for scraping in Excel (but in hindsight I shouldn't be surprised).

I didn't follow the connection between modern looming and scraping though?

Anecdotal, but in the last few months my wife went from not wanting me to turn on AP/FSD to now using it for all her drives.

The v13 models feel like a massive improvement.

Going to have to update the tokenizer for DolphinGemma

https://blog.google/technology/ai/dolphingemma/

How did you fit 550MB of bytecode into the context window? Was this using 2.5 in an agentic framework? (i.e. repeated model calls and tool usage)

There are some options out there, depending on what type of task you're trying to fine tune. I think RL finetuning for DeepSeek e.g. isn't well developed yet, but you can finetune a small LLama model (~3B params) for classification or extraction tasks and it works really well. What sort of tasks were you looking at finetuning for?

same

Obtaining a detailed map doesn't seem difficult. This is the same company that created Street View. E.g. SF has ~1000 miles of streets [1], that would only take 50 hours of paying for manual driving to map all of SF at 20mph.

1. https://data.sfgov.org/City-Infrastructure/Miles-Of-Streets/...

I've been running off a Raspberry Pi for years and it's been very stable. The main things for me were to use a good SD card, to use Ethernet instead of Wifi, and to also reboot the Pi daily via an automation.

I might be missing something but it seems like most of the speed up is from quantization which is commonly used already, and the CPU instance used here isn't that much cheaper (~10-15%?) than a GPU instance that could run the model. For high utilization workloads the extra throughput might be useful though.

Any resources you'd recommend for someone trying to get back into the scene after a long hiatus? Looking for nuanced/informative discussions without too much hype.

Underrated comment.

This sounds similar to "knowledge distillation" in machine learning literature, where you take a large model and use it to train a smaller model. Could you comment on the similarities and differences?

Just for fun, I built an app for magicians to reveal a playing card to a spectator. The spectator is asked to think of a card and then place their finger on the magician's screen - after a few seconds, their selected card magically materializes into view. So far I've been trying it on family and friends and it's been getting great reactions. Disclaimer: you probably would want to have a little bit of card trick experience for the app to be useful.

Apple: https://apps.apple.com/us/app/card-mind-reader/id1548145803

Google: https://play.google.com/store/apps/details?id=com.carddispla...

I'm somewhat surprised you were able to snag this username! Aside from the pop star, it reminded me of Britney Spears' Guide to Semiconductor Physics[0], which was super helpful back when I was in college. Any connection?

[0] http://britneyspears.ac/lasers.htm