Harris (@harris_p10) — posts, photos & videos | PeekX

Harris

@harris_p10

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.

New York harris.nyc Joined April 2012

2.2KFollowers

186Following

630Posts

35Media

1.3KLikes

@harris_p10

Discussion below by @PatrickToulme is from a quarterly meetup of top AI people in New York. Worth reading.
Few things to add.

3. I've grown to really like @bot specially now that it uses best of the available models including opus 5.5. It makes me operate as if I am running a big team of highly specialized employees. It could become the top personal assistant for knowledge work. Never doubt King Elon.
5. Recently shared my thoughts on how human code review is going away soon and how inside of Meta this is already happening at scale. It seemed a bit controversial. Good to see that this is the industry wide trend
6. Yes, fable 5.5 will be the best model ever.. for about 3 weeks before Bel by OpenAI or some other model beats it. Intelligence alone doesn't matter anymore.

Patrick C Toulme@PatrickToulme · 14h

I was at an AI dinner last night in NYC.
Some interesting topics discussed

1. Performance engineers now spend most of their time managing agents and RSI loops. Most perf opts at major inference providers (together, fireworks) are found by agents.
2. RLing too much on code and math has negative effects on your model especially it’s language.
3. People were impressed by @bot and the overall acquisition of @cursor_ai by @SpaceXAI. xAI is a legitimate frontier lab right now.
4. Most people have agents in loop writing, reviewing, and submitting PRs 24/7
5. Most agreed code review is not super productive anymore
6. Consensus was Fable 5.5 will be stunning but OpenAI will beat it in December with Bel

Open quoted post →

4 0 0 8 996 4

@harris_p10

The panic sell-off over OpenAI revenue news is completely misguided.
The discrepancy is from investors trying to compare OpenAI with Anthropic despite differences in how they report partner revenue (see table below):

- Anthropic records the full customer payment as revenue, then records the platform’s cut as an expense.
- OpenAI, for certain partner sales, records only the share it receives.
- The same article itself reports $20B annualized growth going from $30B to $50B in 2 months. That's an insane pace of growth and matters more.
- IMO, this is a big buying opportunity, will share more later.

*Walter Bloomberg@DeItaone · Oct 8

OpenAI annualised revenues $20bn less than previously signalled

Open quoted post →

1 1 1 6 1.1K 4

@harris_p10 reposted

@Altimor

It's irresponsible to let humans review code

Harris@harris_p10 · Oct 7

There's been a lot of debate about AI vs human code review recently. Here's some actual data and observations from inside Meta:

- Engineers now push ~3x the code they did a year ago. Human review capacity obviously didn't triple.
- In an analysis involving 535k+ reviewed diffs, the production-incident rate for AI-reviewed diffs was 1/50th that of human reviewed.
- The revert rate for AI reviewed diffs was 1/3rd of the human reviewed.
- Median review wall time dropped by 35% compared to human-reviewed diffs.
- Experts then looked at every production incident caused by AI-approved diffs. A human reviewer wouldn't have caught any of those issues anyway in a manual review.
- Meta's internal coding agent lands ~3,000 diffs a week, and 59% need zero human edits.

Human review as default for most code changes is already obsolete, and might even be riskier.
Source in comment below.

Open quoted post →

7 6 2 77 11.1K 12

@harris_p10

There's been a lot of debate about AI vs human code review recently. Here's some actual data and observations from inside Meta:

- Engineers now push ~3x the code they did a year ago. Human review capacity obviously didn't triple.
- In an analysis involving 535k+ reviewed diffs, the production-incident rate for AI-reviewed diffs was 1/50th that of human reviewed.
- The revert rate for AI reviewed diffs was 1/3rd of the human reviewed.
- Median review wall time dropped by 35% compared to human-reviewed diffs.
- Experts then looked at every production incident caused by AI-approved diffs. A human reviewer wouldn't have caught any of those issues anyway in a manual review.
- Meta's internal coding agent lands ~3,000 diffs a week, and 59% need zero human edits.

Human review as default for most code changes is already obsolete, and might even be riskier.
Source in comment below.

48 42 22 535 64.5K 401

@harris_p10

Great article on the importance of (ultra) fast inference.
"-Given training phase of a new model is 80% inference (in the form of evals, rejection sampling etc), OpenAI is almost certainly keeping a large chunk of cerebras for its own RSI loop instead of serving it to customers.
- If inference is 80% of the loop and you make it 5× faster, the whole thing gets about 2.8× faster, it's most definitely NOT negligible when you consider the ridiculous race going in between labs to one up their competitors.
"

Beni@beniduboss · Oct 5

Here is my deep dive on why inference speed matters!

It includes everything you should know about regarding what (ultra) fast inference unlocks across agents, research and real-time AI

As promised, I also shared part of my bullish case for Cerebras! x.com/i/article/2106…

Open quoted post →

2 1 0 11 1.3K 4

@harris_p10

Codex: ship something each day for 28 days.
Better challenge: delete something every day for 28 days.
Deleting without making the product worse is harder.
I’d start with the model picker. Wonder what others think.

Tibo@thsottiaux · Oct 4

Over the next 28 days, each day we’ll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset. Let the improvements begin.

Open quoted post →

5 0 0 11 1.2K 3

@harris_p10

Nice bounce but still early days for what $CBRS unlocks. Ultrafast inference is a competitive advantage in itself and Jane street deal is the proof.

Some obvious applications ultrafast is about to transform.
- computer use: every click is a model call and at today's speeds you have to watch it think after every small step. Fast enough and it'll just feel like it's working.
- voice agents/companion apps: having to wait 5-10 seconds in voice mode makes it unusable. Current workaround is to trade off intelligence for speed. Podium, another ultrafast customer is already runs it in their voice stack.
- putting LLM inference in the online path for ads and other rec systems. Think google ads personalized in real time.

There's many others. Jane street literally said: "speed (from cerebras) enables different ways of using the models" and they would not be the last customers.

Harris@harris_p10 · Oct 3

There's so much noise on the $CBRS drop this week. I think it's very misguided and this might be the best buy opportunity in a while:
- The exec sales this week don't mean much at all. They were planned months ago. And the founders are getting some liquidity AFTER A DECADE. Not a big deal.
- There were 19M shares unlocked the same day as SemiAnalysis Nvidia report. Just lots of sell pressure this week.
- As assistants such as muse/dots and voice mode become more common, faster inference will be ever more important even if gains over Nvidia are not 2.5x. Even 1.5x faster decode = 50% more thinking before it starts talking.
- There's no change to the OpenAI deal. Still 750MW, $20B+, through 2028. OpenAI also lent them $1B and holds warrants for ~10% of the company. Not a relationship you walk away from over one launch on Nvidia.

I personally have a big buy order queued for Monday morning.

Open quoted post →

0 4 1 30 3.5K 16

@harris_p10 reposted

Patrick C Toulme

@PatrickToulme

The opportunity cost of not burning your quota / weekly limits / tokens has IMO never been higher.

If you have access to any amount of tokens - burn them. The compute is not saved. Compute is static. It has to be powered and existing 24/7 to serve demand.

4 6 0 39 3.6K 12

@harris_p10

There's so much noise on the $CBRS drop this week. I think it's very misguided and this might be the best buy opportunity in a while:
- The exec sales this week don't mean much at all. They were planned months ago. And the founders are getting some liquidity AFTER A DECADE. Not a big deal.
- There were 19M shares unlocked the same day as SemiAnalysis Nvidia report. Just lots of sell pressure this week.
- As assistants such as muse/dots and voice mode become more common, faster inference will be ever more important even if gains over Nvidia are not 2.5x. Even 1.5x faster decode = 50% more thinking before it starts talking.
- There's no change to the OpenAI deal. Still 750MW, $20B+, through 2028. OpenAI also lent them $1B and holds warrants for ~10% of the company. Not a relationship you walk away from over one launch on Nvidia.

I personally have a big buy order queued for Monday morning.

Beni@beniduboss · Oct 3

I think the market is badly mispricing $CBRS

Same basic phenomenon I’ve argued about with Meta in my first X article:

Wall Street is focusing on the wrong variables

Meta is up 20%+ since said article

Next long-form post will be on inference speed and will include why I think Cerebras is trading at a huge discount for reasons that are essentially irrelevant

Open quoted post →

9 5 1 76 10.8K 16

@harris_p10

This is a much bigger step than people might realize. Up until now, most large scale AI deployments are in the digital world.
Muse gadgets unlock integration of AI into the physical world at massive scale.

And sitting at the center of that is an assistant with free access to 2 billion plus daily users.

Alexandr Wang@alexandr_wang · Oct 2

🚨 MUSE ECOSYSTEM ALERT 🚨

today we are announcing Muse Gadgets!

this is an open-source ESP32 firmware and Linux SDK for anyone to make hardware that works with Muse

we are also releasing our own gadget—Muse Home Link—to enable your muse to work with your smart devices (TV, speakers, etc.)

Open quoted post →

3 1 0 26 1.9K 7

@harris_p10

I have a (conspiracy) theory that extremely verbose comments and clearly excessive unit tests in agentic coding exist so the LLM companies can charge for higher token usage.

Lately i've been seeing 90%+ of code volume be tests that add little to no coverage or value, just tokens.

4 0 1 11 1.1K 3

@harris_p10

Real conversation I had at work today:

I’ll ask my agent to email your agent and cc us.

2 0 0 11 636 0

@harris_p10

AI agents like muse can already make phone calls.
The natural next step is to train them on your voice so they also sound like you. Technically already possible.
In some sense then, that’ll not just be your agent. It’ll be you.

(Also, you can finally call your mom every day)

0 1 1 6 1.3K 1

@harris_p10

Congrats on the big launch! The US now essentially has 4 frontier labs which is really exciting as a consumer. 🔥

And given google's resources this will further lower the cost of using AI in most workflows.

Patrick C Toulme@PatrickToulme · Sep 30

I'm super excited about Gemini 4 Argon. This was a massive effort across so many amazing teams.

Open quoted post →

3 0 0 17 2.5K 3

@harris_p10

dots: can i copy your homework?
@Muse: yeah just change something

dots: $200/month

2 1 0 29 1.5K 3

@harris_p10

dots: it's just like @Muse but instead of free you pay $200 a month.

Stella Kaval@stellakaval · Sep 29

Open quoted post →

11 10 4 162 8.8K 9

@harris_p10

True. Coding models are already commoditized.

If the real "super-intelligence" does come soon, labs will make more money by *using those models* to build and own other industries like biotech instead of selling access.

Patrick C Toulme@PatrickToulme · Sep 29

Hot take - there will be no money in selling coding models at some point in the single digit year future. Coding models will become commoditized and heavily subsidized by larger companies like Meta and SpaceX. Coding is inherently verifiable and how to RL for coding models is already a widely known industry wide recipe.

Anthropic making a wet lab in the Bay Area is a key sign. With coding becoming commoditized, pivoting into AI biology and drug research is the next best thing. Claude discovering a new drug can generate much longer lasting revenue than competing in the coding price war.

Open quoted post →

1 0 0 11 944 3

@harris_p10

100% agreed. My bet: In 2 years reviewing python code line by line would be like reviewing assembly today.
Yes, agents can produce buggy code but so can compilers. Yet we don't go around routinely inspecting assembly.
In fact, in many teams in big tech, it's already becoming exceedingly rare.

DHH@dhh · Sep 28

It takes a while to build the confidence, but there's no future where you're manually reviewing every line of agent code. Not much acceleration in that. You need adversarial agent reviews, you need automated testing, and maybe you spot check. That's it. From prompt to production!

Open quoted post →

3 0 0 9 1.5K 3

@harris_p10

Peter Thiel used to say the best companies to invest in are always monopolies. Competition eats margins. Google & FB were good examples.

Then there’s $SPCX . Good luck building anything to compete with this. It’ll be the most valuable company on earth sooner than most realize.

Geiger Capital@Geiger_Capital · Sep 28

You’re bearish? Starship is shooting out satellites while in orbit and you’re bearish?

0:15

Open quoted post →

0 0 0 2 674 3

@harris_p10

Code review has completely changed in big tech (Instagram) in the last 12 months because of AI:

- I predict human code review will be almost non-existent within a year. Below is how it’s already changed.

- AI agents do a first pass review now. Sometimes, they fully approve to ship.

- Often a human does an AI assisted second pass. But focus is more on high level intent and system context rather than line by line correctness check, which at current code volumes is impractical.

- The agent review is surprisingly good. I’ve seen it catch subtle bugs and edge cases even strong engineers would miss.

- Product teams in most big tech companies heavily A/B test consumer facing changes on small populations and monitor hundreds of metrics to detect regressions.

- So the product quality and user experience wouldn't suffer. It'll likely get better.

6 1 0 10 873 5