I used to only engineer software; then I learned you can engineer everything. Building AI agents to automate supply chains. Cofounder @ Elm AI | @Cornell CS
NYC/London
Joined June 2012
- Tweets1.4K
- Following595
- Followers465
- Likes480
model intelligence on aa, circa april 2026 with deepseek v4.1 flash from this month thrown in
~6 month old sota intelligence at oom faster speed, and many oom cheaper cost
(caveat: yes, not the most fair comparison due to "big model smell" being missing, but directionally true)
While it is true that you can always convert more intelligence into more revenue, I feel like people underestimate the need for Pareto-optimal models that make the right trade-offs between cost, intelligence, and speed.
Over the past many weeks my token usage has been increasing exponentially but so has my model diversity.
Today there are so many use cases where throwing more intelligence at it simply does not improve the user experience or solve the problem in a better way.
everyday there are surprises at work
i measure how unpleasant a surprise is by how much i laugh when the news is delivered to me
cancelled wisprflow after being a long time user, only because today @meetgranola introduced the same functionality, and it feels 100% the same.
the switching cost was about 20 seconds, and I like that the same dictionary that was used for meeting transcriptions is being used everywhere.
this is generally reasonable advice (I usually recommend founders to take money and not be overzealous about dilution for example), but this can backfire.
I personally know cases where investors added negative value to the degree that they ended up costing the company money.
Always backchannel investors' track record and, most importantly, sign standard documents.
ai agents actually still aren't very good, and it's all because of poor engineering
dots is slow and laggy, and not very fun to use
muse is a joy to use, but i can feel the models intelligence limits at times
my personal agents requires me to do work for infra upkeep (craft agents, hermes)
instinct is limited by it's texting only interface (and it ghosted me randomly)
poke was not very bright either, i still keep it though because it's cheap and helps me keep on top of email (I will probably be cancelling soon)
the perfect agent experience today is probably muse + astra intelligence
(all of these are for work + personal work)
just turned on a 24/7 ai agent monitoring our systems after the opencode incident
i always suspected we are constantly under attack by ai swarms, but now seeing the frequency in real time is a bit scary
when we were fund raising, we did the opposite of this, we stated an arr number that was lower than our total "run rate" since we excluded paid pilots, which is not real arr
it was a pleasant surprise for investors to learn that:
- there was no "stretching the truth",
- we were being clear headed about revenue (we broke up arr vs non-recurring)
yeah, our arr number could look a little bit better if I did the "stretching the truth" but imo we earned way more trust by not doing it.
I think the real reason one must resist the urge to mock mathematicians because mocking has never resulted in common ground.
The right thing is understanding their concern, and talking through why they feel a certain way. Everyone will come out with a new perspective.
I am especially interested in talking to mathematicians to understand their point of view.
very poor analysis and penalizes anyone who makes api pricing closer to subs
this measure in fact has nothing to do with "value" in the sense of work value.
we need to actually remove the gap between api pricing and subscriptions as much as possible
incredibly annoying on @ChatGPT -- I figured this doesn't happen anymore? I wonder if I have some misconfigured compaction settings?
Conducting our first true company offsite. While most of the team has met each other, this will be the first time we all get together in person.
Anyone has literature/tips on mistakes we should avoid?
cosign!
i am still looking to fill this role
I’m a YC alum, building a generational company. You’ll work with me daily and have fun while building this company
If you love being scrappy, want to figure things out at the earliest stage, and always curious about how hardware teams build (our customers).
dm me, i’m reviewing apps over the weekend.
There's a new version of this post
elm ai london office is in shoreditch (right by the corgi cafe)
elm ai nyc office is in nomad
when I was raising, many founders helped me, allowed me to work from their space and made intros when I needed it
I want to pay it forward -- if you are in that position, reach out, I'd love to help (and swing by the office to co-work or grab a coffee!)
This is actually a good thing in the medium term. Almost no technology has collapsed in price like AI "intelligence".
Cost per task keeps going down. Yet there is this hedonistic token treadmill pursued in a very wanton manner.
It's obvious that you'd keep getting more value for $200/m (it's been crazy the past year).
And it's an even better thing that the gap between API pricing and subscription pricing is shrinking. No weird restrictions on how we can use out intelligence (like claude code rules, and even to some degree, codex rules).
Not all tokens are built equal, and more importantly, we should be focusing on cost per task. That's the only way we can create some actual gdp, and help with ai diffusion.
Made an ai slop graph (directionally correct, I think the real ratio is 200:1 tbh)
Hi,
Tomorrow we are re-opening the Pro $200 subscriptions to new subscribers, but together with it we are also changing how we calculate the usage for it. In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan.
Now that it's said, let me explain why this is happening and why you will still get more work done than if you were on the Pro $200 subscription one month ago.
(a) We didn't want to compromise in other ways and are committing to not reintroducing the 5h limit, so that you can fully use the weekly usage when you want.
(b) On the subscription, we guarantee that over time you always get more work done and with an increasing level of quality. This means that you will continue to get more value per dollar spent as a result of models getting more efficient and us passing down the improvements in the form of API price reductions.
(c) We don't want to put an incentive on ourselves to artificially inflate the API list prices to make it look like you are getting a lot (and workaround it through discounts, etc). Instead we want to continue to both rapidly reduce prices and increase capabilities of models on the API. This week we introduced GPT-6 Sol and GPT-6 Luna at 50% of their previous price. Over time, we see prices go low enough that it makes sense for most to buy usage as needed without there being a significant gap between what you get in a subscription and what you get in the API for a dollar spent.
(d) Tomorrow, we are adding more things to the subscription that won't draw on the usage, I won't reveal what that is yet.
I wanted to be transparent before all the big announcements tomorrow. Lots of new exciting things are coming to the subscriptions that will make it super compelling, but I wanted to make sure to share this change ahead of time so you can all understand it before we shower you with good news.
Codexingly,
Tibo
nobody can predict the future -- this is from an oai researcher in 2024
"i really don't see how <labs> secure the trillion-dollar clusters needed to get to GPT-5+ capabilities rn"
today, assuming that the AA score of 23^ is the gpt-5 score (flawed by directionally okay), here are the labs that have gpt-5+ class models:
Anthropic: Claude Opus 5.5 (max), 58
OpenAI: GPT-6 Astra (max), 53
Meta: Muse Spark 1.3 (max), 48
SpaceXAI: Grok 4.7 (xhigh), 46
Xiaomi: MiMo-V2.6-Pro, 46
Alibaba: Qwen3.8 Max, 45
Z AI: GLM-5.3 (max), 45
StepFun: Step 5 Preview, 44
Kimi / Moonshot AI: Kimi K3 (max), 44
Google: Gemini 3.8 Flash (high), 41
DeepSeek: DeepSeek V4.1 Flash (max), 39
Motif Technologies: Motif 3, 34*
Institute of Foundation Models: K2 Horizon 375B A23B, 31
Apodex: Apodex 1.1, 30*
MiniMax: MiniMax-M3, 29
Nex AGI: Nex-N2-Pro, 28*
Upstage: Solar Pro 4, 28*
Thinking Machines: Inkling Small, 28*
China Mobile: JT-4.1 Flash 236B A21B, 27*
Multiverse Computing: Quasar 438B (max), 27
Tencent: Hy3, 25
InclusionAI: Ling 3.0 Flash, 25*
^ yeah, some of these are benchmaxed, but this is to prove a directional point.
