Backend engineer. AI infra, Chips and quantum computing/sensing. I post how the machines actually get built.
Thane, IND
- Tweets 4,647
- Following 519
- Followers 552
- Likes 6,652
Exactly. this is the kind of stars we need in India. We all old and young students will line up all classrooms to be there.
1
66
Met him multiple times in conferences and IISER Kolkata. He is always warm and feels like a colleague.
I will believe it as he put out his neck out and say he has never experienced casteism in IITs.
In my 8 years as PhD student at @iiscbangalore and 17 years as faculty member at @iiserkol, I have personally not seen a SINGLE instance of caste-based differential behaviour being inflicted on a student. I have graduated 13 PhD students and supervised around 40 MS theses. #1/n
151
Meet @jakeadler . He is 21 year old High School Graduate, who started building software at 12, hardware at 17 and chemistry at 19. He has no college degree or any formal education.
He is a Thiel Fellow, building @pilgrimlabs to build a 50 pound portable device ((ARGUS)) to detect biological threats like covid and anthrax. They are already working with CDC for initial deployments.
1
99
It is not easy when you have disagree with Harsha Bhogle. But we are extremely behind the likes of South Africa and Australia. Our field and fitness are extremely dull.
India are way way ahead of any other team in in the women's game in Asia. Now to challenge other countries more frequently. Need one more top middle order batter, even after Jemimah returns.
1
118
ResetSanta is here. I hope @claudeai gets their own santa.
Ladies and gentlemen... start... your... ENGINES. We are almost Tuesday and I promised a reset for Tuesday. Among some other things. See you soon.
1
129
I have been working on three things parallely. The boilerplate being written by the AI helps you do it.
- A portable lfmri digital twin.
- A quantum sensor digital twin.
- A personal graph memory/context extraction tool.
The first two are almost done.
The third one has a REAL problem.
contd...
6
2
81
I have of-course tried everything from honcho to mem0, to see experience them. And seriously, I was disappointed by what I saw.
My tool is half baked for general work but works well for the projects I needed it for. I also added some graphdb with PGgraph(github.com/evokoa/pggraph) as I have a lot of experience with postgres.
18
I am mentioning this last but start it was actually for me the most important part.
Write a verifier. another agent that is hyper-focused on verification/evals. I haven't tried local slm's yet. the future goal when budgets align would be to fine tune models locally (jev and the clones) to see if I can write a verifier.
17
- Force the llm to spit out a "Chain of thought" before doing any thing.
- Force the llm to give examples of negatives, postives, false-negatives, false-positives.
You are not tuning the model. you are tuning your instructions and your own expectations of those instructions.
10
While I failed quiet a few times, the key lessons I learned were,
- You conversations with the chatbots need to be clear and centered around the goal. The agent should know your goal. the longer the chat more difficult it is for YOU to be coherent around the goal. And the models themselves dont like long conversations.
6
- Originally goal was not context extraction. just wanted to know what the model thought our conversation was about.
- Context extraction does not give you this directly. Because it is a non-deterministic setting & b/w generation of chat & generation of summary/extraction there is context drift, hallucination, which leads to generation gap.
30
I understand that jobs and interviews are important and undeniable. But every space I join, every meetup I attend and every person I DM. That's the only topic.
No one talks about the tech. Jev came and went. No discussion. Now we have assistants. They are next bull run in tech but no discussions.
2
55
This. No one counsels the students about what you should expect from the school/teacher/classroom. And then again the same student becomes the teacher and makes his own definition of the classroom.
a lot of students have a hard time with this so let me spell it out:
- class is for asking dumb ass questions to someone that knows the material
- learning happens on your own outside the classroom
- office hours is to ask adversarial questions to the main grader about exams
1
2
144
Dhawal Yogesh Bhanushali retweeted
join us if you want to end AI slop!
tastelabs.com/careers
27
16
529
66,405
On one side these labs are releasing reports of how the the agents did things beyond their control. And now they want to give the same agents access to wet labs :/
Anthropic has opened a wet lab in the Bay Area, extending its life-sciences work from simulations into physical biological experiments. The company's goal is for Claude to run laboratory robots with minimal human involvement, aimed at rare and so-called undruggable diseases. Anthropic says its focus will stay on preclinical research rather than clinical trials, keeping it out of direct competition with drugmakers.
76
