aaron (@awlevin) header
aaron (@awlevin) profile picture

something new 👀 prev ai eng @vellum_ai, 1st hire & CTO @thinkfiveable (0 to 1M MAU, consumer edtech), distance runner

Joined July 2010
Photo by aaron (@awlevin)Photo by aaron (@awlevin)Photo by aaron (@awlevin)Photo by aaron (@awlevin)Photo by aaron (@awlevin)Photo by aaron (@awlevin)Photo by aaron (@awlevin)Photo by aaron (@awlevin)Photo by aaron (@awlevin)Photo by aaron (@awlevin)
Pinned Tweet
i built computer use using @typesafeai ! it is 155x cheaper than opus 5, ~20x faster, and generalizes across OS's more on how it works in the vid & thread below:
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x cheaper (w/ output tokens free) • Frontier composable intelligence optimized for decisions AFAICT the shortest path to AI-based economic revolution
70
118
31
2,335
296,220
the outages could be written off as non-cyberattacks because we use compromised AI to diagnose what happened... the AI itself doesn't have to be compromised. only the steps that it's retracing
1
74
aaron retweeted
It's great to finally share what we've been up to! Devin Bots feels like one of the first continuations of the "human-like" feel from Poke, but for engineering. Internally, we treat the Devin Bots like we do co-workers, letting them proactively do work and even merge PRs :)
Meet Devin Bots. From writing firmware to merging thousands of PRs a week, a Devin Bot leads a team of agents to get the job done.
24
11
1
179
10,594
aaron retweeted
our ML team is so studious ☺️
7
3
345
26,650
Maybe hot take but i dont think RSI will be the silver bullet everyone thinks it is. RLVR led to models that are amazing at coding and absolutely awful at vibes. Why would RSI be any better?
Replying to @sspencer_smb @ssi
with the labs pretty close to RSI seems like it shouldn’t be long now
1
99
I’m excited and hopeful for these guys. But, with every new iPhone competitor, I can’t help but feel the marketing misses a very consumer-centric “Steve jobs” touch. Let me explain. First iPod: “1,000 songs in your pocket” NOT: “holds 2gb of songs, plays mp3, syncs across devices” Just “1,000 songs in your pocket” iPhone: “a phone, an iPod, an internet browser” NOT: “it can make calls, hold photos, songs, contacts, wifi or cellular” Personal intelligence shouldn’t be “it has all your context and can manage your whole life for you!” We’re missing the catchy, extremely consumer friendly and easy to conceptualize 1-line catchphrase. Again, no discredit to these guys. Honestly appeasing Twitter can be different than appeasing a general consumer audience, so maybe that’s what we’re seeing here. But maybe that’s also part of the problem. Back in the day, we had the big “keynote” event. That was mostly it for launch day. No channel-specific collateral. Just the keynote and everyone else could go take clips of it and spin it to their liking. Tbf things are so sensationalist these days that maybe the simple 1-line pitches im looking for just undersell it now.
We switched 10 people from iPhone to Blue. Weeks later, none of them want their iPhone back. Introducing Blue: a phone you just talk to. Tell it what you need, and it gets done. It has all the context, can use every single app, and even triages your notifications. We built it for people who always have a little too much to do. The ones buried in messages, notifications, and to-dos. If that's you, get yours today. Backed by Y Combinator + 70 other VCs & angels. heyblue.com
2
196
[read about what’s happening with AI] ME: Wow we’re actually legit fucked this is so bad. [go outside] ME: Wait those people are all acting like nothing unusual is happening. Maybe I was just in a paranoid bubble? Yeah it must be that. Phew. [read more about what’s happening with AI] ME: Wait no we ARE fucked. Shit. It’s all those people outside who are in a delusional bubble. [go outside] ME: Wait...
373
1,008
183
14,599
723,509
Like if you trained this shit on our source code it would get worse at programming. I’m not even sure if Anthropic actually have stricter data policies when you use an enterprise plan paying API prices. It’s wild
7
14
7
1,613
82,756
how are my embedded systems friends doing with AI these days? must be life-changing no longer needing to manually dig through datasheets by hand
76
Replying to @brian_lovin
No way man! Side projects aren’t interesting because they were hard. They’re interesting because they’re like seeing into a person’s day dreams. Don’t ever be embarrassed!
3
5
238
13,692
y'all would be SHOCKED how little information you lose when you ask claude to keep *every* response under 1-2 sentences
1
2
119
I have friends that are skeptical about AI driving down the cost of healthcare. I always refute with “Insurance cos can’t hide behind the fine print anymore; it will make them more competitive and drive down prices” Hope this is true. I’d love to know what nuance I’m missing.
Take the PDF of all your healthcare benefits contracts, including your broker or consultant, put it in your LLM and prompt each one with "Where are they fucking me over and what can I do about it " Take the output. Put the same pdfs into another frontier LLM, take the output from the first, and prompt " What did this analysis get right and wrong and what would you change " I do this across Claude,chatgpt, Gemini and grokbot. Till they mostly agree. They are better than your broker or consultant and can save you money and make sure the contracts are fair. Let us know what happens
1
1
123
In races called, 17 senators are being replaced, already the biggest turnover in the chamber since 1980. The departing senators average age is 70, their replacements 55. The number of senators under 50 nearly doubles, from 8 to 15.
37
1,389
75
15,031
170,525
Replying to @nicbstme
Could more tokens have prevented this meeting?
1
1
3
124
Airport lounges are one of the last remaining things that leave me feeling fully respected as a consumer. It’s only a matter of time.
1
2
156
Hot take: we should be okay with slop. Many use-cases are one-off and don’t warrant any more investment of time or energy than a one-shot yolo
1
100
Games with ELO/numeric rankings are so interesting. You can change user behavior dramatically by simply changing the amount of points rewarded for different interactions. For example, in Chess you could encourage more social interaction by rewarding wins after a rematch
63
today ~ 30K+ people gave up their entire Google Drive for a free pizza each. rly elegant experiment in how we price our personal data 🍕
14
1
1
45
4,665
revisited Duolingo before going on a trip: - quickly hit free tier limits - spent too long learning 5 basic words - couldn't find grammar lessons - skipped ahead and failed the lesson instead, i one-shotted a lesson plan app that: 1. cost less than $2 with opus 5.5 and @ElevenLabs 2. i can use it as much as i want 3. has audio for ~900 words and phrases ($0.30 !!) 4. leverages my existing knowledge of spanish duolingo also had me constantly watching ads and wasting time opening XP boxes. so annoying. this is also accessible to my personal agents because, well, the software and data is all mine i love living in the future so much
1
1
2
312
idk if this'll work for anyone because it's just the dev plan for Clerk but here it is italian.aaronideas.com/ github.com/awlevin/italian
32