Respice Finem.
San Francisco, CA
Joined September 2025
- Tweets128
- Following537
- Followers40
- Likes1.7K
Every week we have had a new model, either open or closed, trying to take on each other. And the sentiment has become Google is left behind. I would say Google is the best model producer, it has the best research, the best scale, most experience and the best data. All the things a model need.
Div Agarwal retweeted
This is incredible news, especially for the AI Agent Engineering & Autoresearch researchers. An overview:
SIA (Hebbar et al.): an open-source agent that rewrites both its own harness AND the underlying weights.
Results:
> 56.6% gain on LawBench
> 502% improvement on single-cell RNA denoising
> 91.9% runtime reduction on GPU kernels
Let's host an emergency hackathon at the house this weekend like we did for @karpathy 's Autoresearch! @kunalbhatia91 !
Superintelligence will be built on Self Improvement.
Today @hexoai, we’re excited to release ‘SIA’ - an open-source Self-Improving AI, to achieve any goal through recursive self improvement.
While trying to solve a problem, SIA doesn't just improve it's abilities by updating it's harness, it updates it's own weights as well.
People with lot of green spots on your github contribution chart, do you prefer quality or quantity?
I see so many on device solutions coming up. Like on device LLM? The whole purpose of cloud was to have complex compute done remotely, cheaper and efficiently. What a UNO reverse on business models.
I am participating in the @claudeai Opus 4.7 Hackathon was building stuff related to Voice AI and memory.
But given the crazy amount of security attacks @Bitwarden @vercel etc etc I might want to create something for this.
Right @getdelve ?
Apart from LLMs, people need to start taking a look into security, especially cryptography and PQC.
No one is going to boast about they made a quantum machine that broke an existing cryptographic protocol or they have the information stored waiting for an advanced machine that can break your encryption.
We need to transition to post quantum cryptography faster.
autoresearch by @karpathy is game changer. I used the idea at Auto Research hackathon @AGIHouseSF for Kaggle competitions to improve models iteratively with agent in loop.
The goal is to have agents do testing and trials for hyper-parameter tuning, feature engineering and more, given you have enough resources/GPUs.
Hopefully we get some unique solutions to some very difficult problems.
Shoutout to @hyperbolic_labs for making this happen.
hiring people for ads company.
I need you to have software skills and should know how to use AI.
not for vibe coders.
in-person, in SF
we build at high pace with people with high endurance.
Dm.
A 0.1 version upgrade in an LLM model should not be hyped this much. Something is wrong. @OpenAI @AnthropicAI