TonyP

Entropy enjoyer

[TonyP @tonypepperoniii]

Made tons of optimizations (accumulated gradients are pretty great) and now can do 16x volumes with lots of layers. I trained one model for a few hours and got some *really* nice results!

I might write a mod so I can start generating stuff in game.

[TonyP @tonypepperoniii]

So, I’ll just not be an engineer. I’m rejoining the software company I founded back in high school as CEO. It’s been slowly growing over the last 12 years (wow I’m old) but I have lots of fun ideas for it :)

[TonyP @tonypepperoniii]

Last week I quit. No more big tech or AI for me. I had fun but I realized I don’t work well in big corporate structures…

[TonyP @tonypepperoniii]

Effectiveness in big tech depends on how well you can get people to coordinate. But as a random eng it’s pretty hard to do that. And my entire management chain leaving def didn’t help

[TonyP @tonypepperoniii]

My world knowledge can be compressed down to 8B parameters which feels about right

[TonyP @tonypepperoniii]

I took the MMLU and got 62% which makes me a weak llama model with worse latency

[TonyP @tonypepperoniii]

It was fun to mess with but an evening with it cost me $30 in tokens and $9.11 in lobster plushies

I’m gonna stick with Atlas for now

[TonyP @tonypepperoniii]

I tried openclaw, it’s amazing but not for the reasons they want it to be

I’m sure someday it’ll be practical but it’s very fiddly and expensive rn. But for silly antics it’s incredible

[TonyP @tonypepperoniii]

I see it loose and it found this adorable little guy and commanded me to buy it

[TonyP @tonypepperoniii]

@proofofgail @claudeai So cute! Claude has been helping me with crochet patterns too

[TonyP @tonypepperoniii]

I'm enjoying turning all my recipes into flowcharts. I feel like there should be a uml like thing for chefs, it's way easier to see what can be done in parallel in this format

[TonyP @tonypepperoniii]

It only triggers on links in the output that aren’t in the input

(Although I agree that description text is misleading and way scarier than it needs to be)

[TonyP @tonypepperoniii]

I actually love this feature! It isn’t an anticompetitive thing, it’s a hallucination and prompt injection safeguard

If you put the URL in the input it goes away

[Nikita Bier @nikitabier]

@sama Perhaps it is you who is manipulating your products to your benefit, by putting warnings on every link to a competitor?

[TonyP @tonypepperoniii]

@DavidSHolz 👋

[TonyP @tonypepperoniii]

don’t even get me started on the interpunct in DALL·E

[TonyP @tonypepperoniii]

We all know OpenAI’s naming was cursed but I just noticed GPT-4.1 mini has a space before mini, and o4-mini has a hyphen whyyy

[TonyP @tonypepperoniii]

Some models have weird capitalization like GPT-4.1 mini Audio. And the o series is doing its own thing with o4-mini-deep-research

[TonyP @tonypepperoniii]

@nearcyan I’ll test it! I’m trying to fix my sleep with a bunch of experiments rn, I’ll put this on the list and let you know if it works

[TonyP @tonypepperoniii]

If I’m pitching LW to someone, the melatonin post is what I show them. So much actionable advice, a lot of it counterintuitive, but with good data to back it up

[near @nearcyan]

it is very counter-intuitive, but the best time to take melatonin to drastically move your circadian rhythm forward is ~7 hours before bed, *not* right before bed! this is despite the fact that the half life of melatonin in your blood stream is only ~40 minutes

[TonyP @tonypepperoniii]

I love how my team has high turnover because everyone keeps leaving to found their own startups and get hired by top AI labs.