Founder of AI Tinkerers. Investing in ambitious AI companies at inception stages. MIT, UWCSE.

USA+
When I was 12 and my dad sat me down in front of an Apple II computer and said, "We're going to learn to program. First thing to know, you can't just say, 'Make me a race car game.'" Then he made me type that in just to see it not work.
74
130
9,104
436,074
Joe Heitzeberg retweeted
fun talk today! every single agent in the world will have their own email soon. it’s the future. email is the perfect communication platform for agents. it’s a 50+ yr old protocol & it’s not going anywhere. give your agent an email in two minutes w/ @primitive📧
30
2
274
10,337
Joe Heitzeberg retweeted
Thank you AI Yoda @JensenHuang for stepping up again AI-yatollah Dario reminds me of Iran's theocracy: the worldview that people need the "guardianship of a jurist". Whether an Islamist or a technologist jurist, we don't need autocratic guardians dictating their values and wisdom on others. Didn't work out well in Iran, and won't work out well here.
THE ONLY REASON Antropic and OpenAI exists is because of the work of this man. What a breath of fresh air. Listen to @JensenHuang leave these hosts speechless as he deconstructed AI fear theater calmly and cogently.
2
3
19
6,391
Joe Heitzeberg retweeted
Excited to announce UFO: a new scalable multiplayer agent orchestrator and memory system. We built from scratch to target pain points we saw with people running agent-first businesses. Give it a try and let me know what you think :)
UFO is live at ufo.ai. A multiplayer agent harness OS you can build your whole company on. Hosted or open source. It's good at code and work.
27
23
466
814,013
In zurich with @ckli85 to learn about his research in computational neuroscience and AI for biomechanically accurate, muscle-actuated models (robots that move like humans, but also potential applications in neuro-recovery)
1
4
390
Lunch in Zurich with some AI Tinkerers
28
805
waymo training in london
3
10
512
ai tinkerers finance and AI leaders in london — alicca bank, london stock exchange, monzo and more with @arizeai
2
9
558
with @arizeai in london with AI & Finance leaders - what’s working? what’s bottlenecked? great insights tonight!
5
646
Stop clanker discrimination now! (Tried to use Muse to get a better fare on Delta, and Delta locked me out of my account because of it.)
5
1
17
1,069
with @james406 in his home turf (Cambridge UK) with 12 AI PhDs and a lot of good food… also this man is insanely ambitious, chill, steadfast abd long term. Watch this space. @posthog. you will be surprised.
7
387
Joe Heitzeberg retweeted
Expect more reactions like this to Muse and other shopping / negotiating AI agents. Reminds me of an investment we made in YAPTA almost 20 years ago - stood for Your Amazing Personal Travel Agent. They guaranteed the lowest price on any flight you booked by constantly checking the price changes over time and alerting you so you could rebook at the lower price. It worked until the airlines noticed what was happening and started raising the change fees and imposing other restrictions making it harder to realize any savings. Today a friend emailed how Muse helped him save $5000 a year on car insurance. Some incumbents will push back but it’s a great time for customers and companies who can figure out how to experiment and take advantage of the new new thing.
8
5
39
4,926
Tonight in Cambridge, UK - researchers from Google DeepMind, MediaTek Research, Cambridge, Imperial, and Oxford, working across diffusion models, reinforcement learning, causal ML, graph neural networks, quantum information, NLP evaluation, and AI-accelerated physical simulation, plus @james406 and I for a private dinner exchange of ideas
6
336
despite evals, engineering best practices and frontier models, nothing will overcome the complexities and nuances of timezone handling.
5
404
2028: "We wanted the humans to feel useful so we gave them a few inconsequential trivial things to deal with”
1
356
How? Inference log analysis. I ask my agent to understand Jev and to look for opportunities, ranked by total tokens and model cost. Then I have it prototype a Jev-based implementation and run it on inputs from my logs and compare outputs from the existing implementation vs. the jev-based prototype, evaluating with a frontier model. [5 of 6]
1
1
103
I've run /goal on these with instructions to iterate and improve until quality is at parity. [6 of 6]
67
Found $1200 / month inference savings today so far. Not just by replacing some LLM workflows with Jev based steps, but by eliminating a lot of context bloat. [1 of 6]
1
4
639
I've found and deployed changed today that will save me around $1200 / month and I'm not done finding mode [4 of 6]
1
57
The biggest savings are from taking simple matching operations and moving them to Jev instead - 60% to 80% cost savings for a few things [3 of 6]
1
1
112
Context bloat: prompts that have a ton of unnecessary context that can be safely removed without impacting quality. This is a good thing to have an agent scan your code looking for! This alone will save $ [2 of ]
1
2
128