Member of Technical Staff at CV
San Francisco Bay Area
Joined July 2022
- Tweets158
- Following238
- Followers125
- Likes273
Adam Blumenfeld retweeted
The Claude Code hackathon is back for Opus 4.7.
Join builders from around the world for a week with the Claude Code team in the room, with a prize pool of $100K in API credits.
Apply by Sunday: cerebralvalley.ai/e/built-wi…
Adam Blumenfeld retweeted
The third stop on our Asia Tour with @GoogleDeepMind
Seoul 🇰🇷
A legendary city for a legendary hackathon - all for Gemini 3.
$100K+ in API credits.
30m call w/ the founders of AI Future Fund
And a member of the 22nd National Assembly
This one made history 👇
Adam Blumenfeld retweeted
The second incredible stop on our Asia Tour with
@GoogleDeepMind
Tokyo 🇯🇵
A special event in a special city - all for Gemini 3.
@supercell collectible figurines for our 3 winners
$150K in API credits, and all to play for...
Catch the action👇
Adam Blumenfeld retweeted
Weekend well spent. The travel was worth it.
Heading back with a lot of love, learnings, and direction.
Thank you, team @cerebral_valley, @joshuajerinn, @AkritiKeswani, and Adam.
Adam Blumenfeld retweeted
I turned a dream that would’ve taken me 40 years to even start into a reality for my 21 y/o self in 3 months.
A thread on @elsewhere_today from idea to launch ⬇️
Adam Blumenfeld retweeted
When I moved to SF I was fed up with how every cafe was closed by 5
So we're building a place for the dreamers & doers after dark
Those who don't fit the conventional 9-5
Launching @elsewhere_today's first popup with @cynthwangg on July 24
Invite below (spots filling up fast)
Adam Blumenfeld retweeted
🥉 3rd Place (Objective Track): PhysicalRL
model trained to create physical representations of CAD using generated STL data
@nullref0 @venkatacrc
Adam Blumenfeld retweeted
My thoughts,
Cursor is good (I pay for it), but it isn't as context aware, performance declines with iteration, and I get inconsistent results in code quality.
Claude Code is the best I've used so far but it's super expensive.
V0 is OK and domain specific which I like, but I think it's better for prototyping and component design than a production system. Their Shadcn importing integration feels half-baked especially as most people use it as a full-frontend tool.
I've been experimenting with fine-tuning SLMs on specific types of tasks, and aim to make something domain-specific and extendable rather than an oracle out of the box. I tried doing that with this problem two years ago (CSXL/solus), when we took it proprietary and tried to ship it with commercial LLMs WE couldn't even pay for our own product.
However, I don't want to be as narrow as something like Replit. I'm mainly trying to build a tool that will help me with my primary areas of development (NextJS, Custom Components with a Component Framework, Unit-tested backend). Through SLMs, I am trying to map multiple inexpensive domain-specific agents to a single task, rather than one language model spinning on the same thing for 10 minutes and leaving a mess everywhere.
Adam Blumenfeld retweeted
We are looking for a student researcher to work on video understanding plus 3D, in Google DeepMind London. DM/Email me or pass it to someone if you feel it may be a good fit!
Adam Blumenfeld retweeted
Nature research paper: Distributed quantum computing across an optical network link
go.nature.com/4aQeiMU
Adam Blumenfeld retweeted
Riff (tariff management and intelligence) is launched. Request early access now.
riff.csxlabs.org/
Adam Blumenfeld retweeted
We have to take the LLMs to school.
When you open any textbook, you'll see three major types of information:
1. Background information / exposition. The meat of the textbook that explains concepts. As you attend over it, your brain is training on that data. This is equivalent to pretraining, where the model is reading the internet and accumulating background knowledge.
2. Worked problems with solutions. These are concrete examples of how an expert solves problems. They are demonstrations to be imitated. This is equivalent to supervised finetuning, where the model is finetuning on "ideal responses" for an Assistant, written by humans.
3. Practice problems. These are prompts to the student, usually without the solution, but always with the final answer. There are usually many, many of these at the end of each chapter. They are prompting the student to learn by trial & error - they have to try a bunch of stuff to get to the right answer. This is equivalent to reinforcement learning.
We've subjected LLMs to a ton of 1 and 2, but 3 is a nascent, emerging frontier. When we're creating datasets for LLMs, it's no different from writing textbooks for them, with these 3 types of data. They have to read, and they have to practice.
Adam Blumenfeld retweeted
the legend @srush_nlp giving a talk on o1-adjacent matters? Immediately registered. 👀
Adam Blumenfeld retweeted
Here’s the record of the message between UCLA and Stanford.
It’s the internet’s birth certificate.
Adam Blumenfeld retweeted
📢Applications are open for the @allen_ai Research Internship at the Aristo team! Join @ai2_aristo to do open research on topics like:
🤖 Agentic planning, reasoning, and learning
📖 Learning, representation, and reasoning
⚙️ Model tuning and specialization
📊 Data-driven discovery
🔬 Automated and AI-assisted scientific discovery
📅 DDL: Please apply by November 30, 2024 for 2025 summer internship.
🔗: job-boards.greenhouse.io/the…
