Dharmesh Kakadia (@dharmeshkakadia) profile picture

Building https://teapaw.org/t.co/VcaMs28aTa to give post-training superpower to everyone. @mixtrainai Past @nuro @zoox @Microsoft @MSFTResearch @iiit_hyderabad

San Francisco, CA
Joined August 2011
Photo by Dharmesh Kakadia (@dharmeshkakadia)Photo by Dharmesh Kakadia (@dharmeshkakadia)Photo by Dharmesh Kakadia (@dharmeshkakadia)Photo by Dharmesh Kakadia (@dharmeshkakadia)Photo by Dharmesh Kakadia (@dharmeshkakadia)Photo by Dharmesh Kakadia (@dharmeshkakadia)Photo by Dharmesh Kakadia (@dharmeshkakadia)Photo by Dharmesh Kakadia (@dharmeshkakadia)Photo by Dharmesh Kakadia (@dharmeshkakadia)
Rex is legit awesome! So many small things and attention to details! Today I discovered that this go to/restore window is much better to maintain my "sessions" views in terminal! 🫡 @mitchellh and the team!
1
374
Post training works, is boring, and profitable.
Wix's Base44 went from near-zero gross margin in Q1 to 58% in Q3 Most of that gain is down to their in-house Base 1 model which they fine-tuned using 10M+ real user interactions on the platform.
2
220
Don't let @dylan522p see this :)
we built an inference engine for deepseek v4.1-flash from scratch. the core engine is just rust + cuda. no pytorch, no sglang, no dsls, etc. ~1740 TPS decode on 4xB300, 2060TPS on 8xB300. 1M token prefill in ~12s. it's up on github
146
Airbnb for compute might be better name that neoclouds :)
Long-running sandboxes are like Airbnbs. You rent the whole place for a month; the owner makes a predictable margin; there's no overhead between guests. Short-running sandboxes are like hotels. - Guests come for one night - front desk staffed 24/7 - amenities maintained constantly Massive overhead per night relative to occupancy. What many sandbox companies do is price for the Airbnb case while selling hotel-style usage.
1
1
100
A lot of people haven't updated there priors about general physical AI capabilities of frontier models. And they are going to be very surprised.
Replying to @claudeai
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
1
31
1,440
Since the all the personal assistant/agents like instinct etc are on iMessage, why not leverage apply cash payment as agentic payment method instead of the new protocols? i.e. just "request $ when you need to buy something on my behalf?
145
Frontier models still seem to make simple math errors, all the time, if you are not looking carefully!
1
2
101
Dharmesh Kakadia retweeted
No one talks about how deeply humiliating starting a company can be
98
82
36
1,174
228,436
Everyone is overestimating how much demand is there for productivity personal assistant. Muse, instinct etc is good, but global reach requires disposable tasks "normal" people can outsource. Most of such tasks already solved with special cases - order me food/taxi or give me news
3
140
Every time someone asks me "Why won't hyperscalers do X ?", I am reminded of this classic article about Sun and incentive structures at play. fullhoffman.com/2006/03/20/t…
1
96
There is a flip side of this. As a manager/leader, your no 2 job is to build momentum for your team(s). No 1 job is to make sure you have the right people and they are working on the right things.
new post: the senior engineer death spiral sunilpai.dev/posts/the-senio… a friend just started a big job and asked for some advice. so I braindumped a monologue about a super common failure mode I see with engineers and posted it here, hope it helps whoever it can.
5
335
A lot of people are going to be surprised about OpenAI's robotics progress and ambition.
4
286
Openrouter's latency and volume data is the most reliable proxy for talents at these companies. Not just absolutes but how it trends over time as well.
Nice stats from @OpenRouter that illustrate how @togethercompute delivers solid and scaled performance for agentic workloads. We are serving GLM 5.3 and GLM 5.3 Flash @ top decile of TPS, latency, cache rate, and doing it at large volumes, 23% and 30% of all OpenRouter traffic, and OpenRouter is a fraction of our overall API traffic. Running scaled inference for agentic applications is a lot more than maxxing a single metric ... you need to optimize on all dimensions and doing so at scale with reliability. openrouter.ai/z-ai/glm-5.3 openrouter.ai/z-ai/glm-5.3-f…
1
189
I think Astra computer use seems to have gone under the radar a bit. Its a very big deal.
1
1
128
In today tale of AI : "I'm not being dumb, but I did keep answering a question you weren't asking."
120
Gemini is again testing side by side responses! New model must be releasing soon!
1
155
I hope Anthropic paid enough to whoever taught Opus the most important AGI skill : "I made that up".
126
Everyone is completely sleeping on multimodal native workflows. We are going to see order of magnitude more demand for compute from multimodal workflows as they come online - video, robotics, bio,..
If you are not coding all day, your personal AI token usage is comically low. Only exception is video generation. A creative can easily spend more tokens than a software engineer. Co-work, customer support, legal, healthcare doesn’t have token-market-fit yet.
196
Why do you need your own model? "We’ve implemented new interventions that limit Claude’s effectiveness for requests targeting frontier LLM development (for example, on building pretraining pipelines, distributed training infrastructure, or ML accelerator design"
Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use. Its capabilities exceed those of any model we’ve ever made generally available.
2
3
151