Harris

ML engineer @Meta. Longtime engineer, newly posting.
Follow for an insider's view of AI: what works, what breaks, what’s hype.
Opinions my own.

[Harris @harris_p10]

Appreciate the shoutout @PatrickToulme.

Quick intro for new followers: I’m a machine learning engineer at Meta where I’ve worked for 8+ years.

Over the past few years I’ve worked on Instagram across ads systems, shopping, growth strategies and most recently people recommendations.

I’ll be posting more about AI and its practical applications at scale, what works, what breaks, and what’s overhyped, drawing on my own experience and learnings. Plus some career stuff.

Based in NYC. Usually playing tennis when I’m offline (hit me up if you play).

Look forward to connecting.

[Patrick C Toulme @PatrickToulme]

Go follow @harris_p10. He works inside Meta and knows whats going on rn

[Harris @harris_p10]

@0xrwu Well sure but Meta would have the second best model

[Harris @harris_p10]

@alexandr_wang Everyone knows the newest scaling law: agent performance scales up linearly with thiccness

[Harris @harris_p10]

@kobelum Partiful getting worried by moves is a bull signal. They wouldn’t care if they thought it was nothing. Keep it up

[Harris @harris_p10]

@kunchenguid "cycles started to synchronize 😂"

meanwhile gemini has been completely on birth control

[Harris @harris_p10]

@PatrickToulme All labs are using internal teacher models that are at least a few months ahead. It would make no sense not to. But more than the smaller and cheaper I am liking its output writing style.
But completely disagree with one of your points
https://t.co/cCRap0hj8K

[Harris @harris_p10]

- Completely disagree with “First RSI” model part. What evidence do you have for it?
- If Anthropic had hit RSI at this level, that would be AI news of the year. They wouldn’t hide it.
- A teacher model was almost certainly used which explains the smaller and cheaper part. That’s not really news at this point.
- There was most likely heavy RL on top after distillation and that almost fully explains the benchmark improvements over fable.

[Harris @harris_p10]

@tonygaorx looking at the output style this seems like a lot of post training

[Harris @harris_p10]

- Completely disagree with “First RSI” model part. What evidence do you have for it?
- If Anthropic had hit RSI at this level, that would be AI news of the year. They wouldn’t hide it.
- A teacher model was almost certainly used which explains the smaller and cheaper part. That’s not really news at this point.
- There was most likely heavy RL on top after distillation and that almost fully explains the benchmark improvements over fable.

[Patrick C Toulme @PatrickToulme]

Opus 5.5 was obviously trained by a much bigger teacher model. Probably Model-2 Mythos. Opus 5.5 is the first model trained from RSI and also distilled from the internal Ant teacher model.

This is how Opus 5.5 is both smaller and cheaper. It is an artifact of teacher model distillation.

[Harris @harris_p10]

@DavidPawlan Hello, I’m interested

[Harris @harris_p10]

@PatrickToulme They did remove the em dashes so maybe AGI is near.

[Claude @claudeai]

Opus 5.5 communicates more naturally, addressing some of the most common feedback we heard on Opus 5.

It puts the most important information up front and follows the writing rules you give it, which makes long sessions easier to follow.

[Harris @harris_p10]

@noself86 @PatrickToulme Every lab is using a teacher model at this point. It will make no sense to not do that

[Harris @harris_p10]

@Johnnyp135DFS Which part do you disagree with?

[Harris @harris_p10]

@jonathanzliu It’s the UNGA week. Not usually this bad

[Harris @harris_p10]

@brian_armstrong Certain AI doomer claims are often also quite contradictory. One essay claims AI is going to cure cancer. The very next one says we should pace it.
If something is going to cure cancer we probably shouldn’t slow it down.

[Harris @harris_p10]

i am actually very bullish on ChatGPT ads because they're presented to the user in a very high intent moment.
("ok, which of these hotels in Budapest would i like the most based on everything you know about me and this trip")
And people have started to blindly trust ChatGPT's recommendations.
also, as it stands muse data isn't used in meta ads systems.

[Harris @harris_p10]

@algoflows only one of these is highly montizable and it't not math.

[Harris @harris_p10]

Looks like this card might be the best kept secret of holding bitcoin:native on Coinbase no one knows about.

4% cash back on every transaction, 5% on travel (in bitcoin) and no yearly fees. I’m not aware of anything close.

[Harris @harris_p10]

@lukaivicev I don’t currently have it but seriously considering Coinbase gold.
4% bitcoin return on every transaction, 5% on travel, no annual fee.
I like my platinum but this return percentage on *every l* transaction is hard to beat.

[Harris @harris_p10]

@davidsven can't wait for Connect 🥳

[Harris @harris_p10]

@tonygaorx Wrote about a big example here. Others are making appointments, reservations so far.

[Harris @harris_p10]

this is a serious issue. i have chatGPT negotiating a longstanding and complicated dispute with a major airline (non US) on my behalf for a few weeks. It handles all the back and forth itself.

this morning i saw my sent emails and was shocked to see how aggressive and threatening it was at times.
problem is, if i don't give it auto email send permission, i become the bottleneck and the automation isn't nearly as useful.
haven't found a good solution to this so far but i am all ears.

[Harris @harris_p10]

This argument has it completely backwards. "People don't want to do stuff" is a bull case for assistants not a bear case:
- personal assistants aren't just for hyper-optimization of daily life. They're also just useful chatbots for learning, researching and simply looking things up. A much better version of search.

- yes most people want to hang with friends, complain and scroll reels. but automating menial tasks lets them do *more* of that not less. booking a plumber appointment in 60 seconds instead of 20 mins means 19 more minutes of TikTok on your couch.

- there's already proven enormous demand for reducing life friction in general. e.g., delivery apps, auto-pay, personal assistants (the human kind) etc. AI personal assistants just take it to the next level.

The AI assistants empower people to do less of what they don't wanna and more of what want to do.

[Taylor Lorenz @TaylorLorenz]

“Absolutely nobody wants to realize their potential. People want to hang out with friends and complain. That’s how the world has worked for ten thousand years.”