Aren

Often skeptical, often optimistic

[Aren @ArenRendell]

@connacher_ You should’ve owned him by saying “I’m full stack.”

[Aren @ArenRendell]

My initial reaction is that Opus 4.8 has better feel than Opus 4.7. Cautiously optimistic.

[Aren @ArenRendell]

@rSanti97 @mattyglesias +1, that tweet sucked. Was trying to figure out if there was something written in the letter to make it make sense and obviously there isn’t😭Would delete asap

[Aren @ArenRendell]

I’ve been worried about AP (accounts payable) phishing for months. I have gotten more worried. LLMs are useful for this. Few people with power to make changes ever think about how invoices get paid.

It is solvable and not entirely unprotected. But no one is thinking about it.

[Aren @ArenRendell]

This is really cool from a person who runs the company that launched ChatGPT, which was an insanely expectation-exceeding product

[Sam Altman @sama]

being a dad is the thing that has most exceeded already-high-expectations in my whole life

[Aren @ArenRendell]

I asked Claude about this and it said “consumers don’t buy tokens directly” because they get the subscription. Can’t believe the low-quality of LLM outputs. Consumers are buying tokens from the API now, too.

[Aren @ArenRendell]

To my knowledge, the CPI does not account for the massive deflation happening in cost/output in LLM tokens. Why shouldn’t it?

[Aren @ArenRendell]

@connacher_ Doing great, thanks for asking.

[Aren @ArenRendell]

@connacher_ I see. Hmm. Maybe just ask Opus 4.7 and GPT 5.5 to talk about it until they figure it out.

[Aren @ArenRendell]

@connacher_ Any theories as to why this is true?

[Aren @ArenRendell]

@connacher_ Layer of LLMs poring over data to remove PII, starting with on-device models and moving to a final smart model API call?

[Aren @ArenRendell]

The obvious issue is getting sufficient liquidity and activity to make the markets useful. There’s also a huge challenge in determining who counts as an insider. Someone may be generally an outsider, but bring a valuable perspective.

But very into the theory. Would be amazing info without revealing specific intel.

[Aren @ArenRendell]

Why didn’t OpenAI slightly undercut Anthropic on API list price?

$4/22.50 or $5/$20 would’ve meaningfully altered decision-making. That process is much shallower than technologists would like to believe.

I feel that 5.5 is cheaper to run over time. But the list price matters!

[Aren @ArenRendell]

Big companies are launching a LOT of things. I’ve been feeling this crowd out adoption of any one of their things. I spent a lot of time talking with firm leaders in the last two months. They largely share my feeling.

[Aren @ArenRendell]

Like humans, models can get better at some things while getting worse at others, it seems.

[benedict @bqbrady]

I am a little surprised 5.5 did not perform better on my benchmarks. See the rest of the charts and read the full post here: https://t.co/Z8E9Sc9Iou

[Aren @ArenRendell]

I don’t want to build this because food is a personal passion and hobby. I don’t currently want to pollute this hobby with my work brain right now.

But for anyone who does: I will share my thoughts and provide feedback freely.

[Aren @ArenRendell]

As a foodie and agent power user, I hope something in MPP/x402 world takes off and solves dinner reservations. I’m skeptical.

The bottleneck is clear: everyone wants to go to the same restaurants, as they should! Flour + water is incredible. Mediocre Italian isn’t.

[Aren @ArenRendell]

The agent version might allow restaurants to join a consortium to “fill last-minute seats,” then use the speed and reasoning abilities of agents to coordinate across parties, payment-enabled diner agents included.