social.nishtahir.com

This took for ever but I finally got Ghost's activity pub integration to work with my blog. So you can now follow and interact with me there

@nish

Learn from my mistake, if you run a #gotosocial instance please backup your DB. I accidentally restarted by containers to debug an issue while migrations were happening and ended up nuking my instance.

Luckily I had weekly VPS backups I could restore from but that was way more difficult than it should have been.

I had some fun doing some reverse engineering and building a debian package in Rust https://nishtahir.com/building-an-ubuntu-service-for-my-antec-flux-display/

Wrote up some thoughts I've been having on the value of expertise. https://nishtahir.com/the-value-of-expertise/

If you've seen talk of ghost engineers, I break down some concerns I had after reading the paper https://nishtahir.com/on-ghost-engineers-and-proof-or-a-lack-thereof/ #programming #softwareengineering

I am now on bluesky https://bsky.app/profile/nishtahir.com

Sometimes personalization can be incredibly weird. Amazon has decided I want a pink T-shirt press. Have I ever interacted with this? Yes, because it recommended it and I wanted to figure out wtf it is. Now it haunts me non stop. When I visit the site, open the app, my notifications...

I sometimes look at weird electronics so I forget what it is and click into it to read the description which continues the chain reaction of Amazon presumably thinking I'm on the fence about buying it. Incidentally, if I buy the damn thing will it go away?

I just spent the last 3 hours debugging an issue only to discover that the backend I'm integrating with works differently if your requests end with or without a trailing '/'.

I updated to iPadOS 18 to get the new calculator app with Math Notes and gave it a go. Since I've been doing more math in service of learning ML I am very qualified to evaluate this.

Overall the most painful part is getting the app to recognize my handwriting. It misses the mark so much that sometimes after using the dial that shifts numbers up/down it doesn't recognize it's own sometimes.

Writing longer equations or equations with multiple terms can be frustrating. Misses an equals here, reads a as aa somewhere etc.

If you manage to get that in place, the graph at least for me couldn't figure out where the center was so I had to scroll manually. Recenter doesn't work for me.

Overall, it's cool and has potential. I hope Apple works out the kinks. For now though, probably better off using Wolfram Alpha or something

I remember a time apple used to make fun of Android for making tablet apps that were "stretched up phone apps". This is the new calculator app's basic view.

Google released a line in their earnings call that is generating a lot of buzz.

> "More than a quarter of all new code at Google is generated by AI, then reviewed and accepted by engineers"[1]

I don't deny that AI has a lot of utility, I work on and use them everyday but I think the blanket 25% of all code doesn't leave a lot of room for nuance.

- What does it mean to get reviewed an accepted by an engineer?
- If AI writes 10 lines of code but the engineer has to rewrite 9, does that count as AI generated code?
- What types of projects did they work on? "An AI wrote 25% of the code for a completely novel system written in an esoteric programming language" would leave a very different impression than an "AI wrote 25% of the code for an expressjs server".

In my experience, AI does some things well - typically the more trivial cases, but struggles to deliver on more complex tasks without human intervention. This doesn't mean that they have no utility, but that there's a learning curve and expectations have to be tempered in terms of productivity gains. It's a win if no one has to write boilerplate, but it's not going to suddenly 10x anyone's output.

I think this is just more fuel for the hype cycle that unfortunately sets up unrealistic expectations for business leaders that latch on expecting 25%+ productivity gains from their teams.

[1]: https://blog.google/inside-google/message-ceo/alphabet-earnings-q3-2024/

Apparently, you can submit health data to Grok for advice. Looking at the tweet. People have been submitting images of x-rays and lab tests (can't say it's a good idea but oh well)

"Ryan Tarzy, CEO of health tech startup Avandra Imaging, says in an email that Musk is trying to speed up Grok’s development by bypassing direct-sourcing the data rather than obtaining the data from a secure network where patient data has been de-identified."[1]

the results seem pretty decent from a lay person's perspective. However some experts weighed in calling out inaccuracies. Please don't take healthcare advice from chat bots or give them your medical information or. Seek out health care professionals.

[1]: https://www.fastcompany.com/91218769/elon-musk-wants-you-to-submit-medical-data-to-his-ai-chatbot

TIL this is the openbsd mascot.

> Subsequent releases used variations such as a police daemon by Ty Semaka,[130] but eventually settled on a pufferfish named Puffy.[131] Since then, Puffy has appeared on OpenBSD promotional material and featured in release songs and artwork.

I've lost count of how many "AI won't replace humans, but humans using AI will replace humans without AI" videos I've seen, which showcase the AI not working while the human ignores the output and just does the work.

#askfedi. I'm trying to step up my fediverse game. I just learned that following hashtags is a thing. What hashtags do you follow that you've found useful?

Micro optimization is one of those things people say never to do. It's supposed to be the last thing on the list but I cannot help that it's literally one of my favorite things. I refactored a script that took 40 minutes to run in 4. It's a script that runs once a week at 2 am that no one else interacts with. But I will know how much time it's saved and isn't that what matters?

Great summary on a paper apple put out on LLM reasoning capabilities[1].

> we found no evidence of formal reasoning in language models …. Their behavior is better explained by sophisticated pattern matching—so fragile, in fact, that changing names can alter results by ~10%!

I've had similar observations when trying to get LLMs to play complex games specifically Pokemon S/V. The problem i observed is that reasoning tapered off with multi step strategy so I always had to provide some form of external decisioning. They worked great for the more simple pattern matching aspects like navigating the UI, but just couldn't play the game unfortunately

[1]: https://garymarcus.substack.com/p/llms-dont-do-formal-reasoning-and

Getting closer. Looks like something that would be at home in Palworld.

This is not exactly what I had in mind but accurate. Touche Google... Touche

I've been using #NotebookLM for a while. As a study tool, I think it does many things well but has a long way to go to be an effective product - for example, it doesn't save chat logs. The podcast generation is currently its best feature.

Here's a of sample I generated based on a paper I've been studying about direct preference optimization - https://arxiv.org/abs/2305.18290