PhD, @MIT. BTech, @iitmadras. AI, ML, Climate, Upskilling, Education. Views are my own.
Bengaluru, India
Joined September 2009
- Tweets3.4K
- Following789
- Followers3K
- Likes5.1K
Deepak Subramani retweeted
Yup, it's the new form of performative busywork.
Deepak Subramani retweeted
I think voter apathy or long weekend is a very small problem compared to the data errors that have a very high probability and plausibility. The attack on hard-working tech bros who may have taken off to a resort after voting at 7 am is uncalled for.
Deepak Subramani retweeted
Everyone wants to have solved a hard problem, but almost nobody wants to live inside the grind of actually solving it. The breakthroughs belong to the few who are obsessed with the process, not just the idea of the outcome
Deepak Subramani retweeted
Finally! This might be the end of Leetcode interviews
Code comprehension is still important, especially in the (early) age of agentic coding
Goodbye to LeetCode? Google just updated its interview process. AI is now a part of it.
They have introduced a new round called Code Comprehension.
The idea behind this round is to drop the candidates into a 200-500-line codebase with Gemini available for use.
This round evaluates:
- Can you read someone else's code
- Are you able to debug independently before utilizing AI
- Can you catch when Gemini is wrong
Deepak Subramani retweeted
Claude Managed Agents dynamic workflows are now available in public beta. It's a new type of multiagent orchestration, built for your most ambitious workloads.
A lead agent writes a plan that runs across many agents in phases, combining the results at the end.
Deepak Subramani retweeted
the most common failure case I see is when people are working outside their domain of expertise and don't know how to be precise with their prompts and plans, so they have to spend a lot of turns iterating imprecisely
Deepak Subramani retweeted
I am surprised that people are surprised this is how I prompt Claude.
Talk to Claude the way you would a coworker. There's no secret to prompting. There's no need to be overly scaffolded or prescriptive for most tasks -- give Claude a goal, and it will figure it out.
Back in the Sonnet 3.5 days, your prompt mattered a lot. Nowadays, it's much more important to communicate to the model:
1. What you want it to do
2. How much effort you want it to spend
3. How it should verify that it did the right thing
Deepak Subramani retweeted
"explain how Shazam works"
I had no idea
Karpathy is so right about this
custom explainer videos will change education
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
Deepak Subramani retweeted
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
Deepak Subramani retweeted
Because so much of the debate about AI and jobs is driven by fears of outright replacement, there's a much more likely way in which AI could make work worse that isn't getting much attention. Here's my case:
1) A growing share of knowledge workers are managers of AI agents rather than doing the work themselves.
2) Being a manager (of humans) actually kinda sucks. You don't get to practice the craft you trained for; being responsible for others' mistakes is super stressful; and your attention gets split between dozens of things.
3) But we managers don't get to complain — it's higher pay and higher status, and mentoring people and watching them succeed is incredibly gratifying. We chose to get promoted.
4) Being a manager of AI brings all of the negatives without any of the positives. When everyone is a manager, it's not high pay / high status. And the role change is forced upon you.
What am I missing?
Deepak Subramani retweeted
Periodic reminder: professors are terrible at forecasting the future. If they were any good at it, they wouldn’t be professors: they’d be making millions in the stock market.
Deepak Subramani retweeted
We hired a CS grad last month who genuinely thought writing clean algorithms from scratch was what got you promoted.
He spent 3 days writing a custom parser. I showed him how to do it in 20 seconds with claude, wrapped in an API, and told him to go hit the gym
He looked like his childhood was ruined.
Computer science degrees are officially liberal arts degrees now. Modern tech is just plumbing.
Deepak Subramani retweeted
I used Opus 5.5 to formally verify the Claude Agent SDK using Lean. A couple short prompts = 16 PRs fixing various bugs and race conditions. Video attached.
TLA+ also works well. I sometimes combine Lean and TLA+ to look for issues around data flow, concurrency, and state mgmt.
I don't know either language well, but Claude is excellent at both. This approach is super useful for formally modeling your code and finding bugs that a human probably wouldn't have spotted.
Is formal verification the future of coding (or at least, bug finding)?
Deepak Subramani retweeted
As organizers of CVPR 2025, we were one of the first to introduce authorship limits with a limit of 25 per author. In light of the excessive amount of submissions of ICLR 2027, many here were proposing stricter authorship limits. I want to offer my 2 cents: I don't think authorship limits are useful to reduce the number of submissions. Authorship limit is a tool to limit people putting their names into too many papers that they may have not contributed enough, which has served its role already with the current limits.
In CVPR 25, more than 90% of the first-authors only have 1 first-authored submission, and more than 99% of the first authors have at most 2 first-authored submissions. With this in mind, rate limits for submissions are likely not going to help reducing the number of submissions much, if at all. The problem is rather that everyone who wants to be an author nowadays can be one with AI-generated papers, not necessarily a few authors spamming the system, and I would assume it is more likely the case right now, although the proportions may have shifted a bit.
Area chair desk rejections (maybe with automatic flagging before manual checking) before the paper going through the reviewing process might be a viable approach right now, and I think it's worth testing that in upcoming conferences. But submission fees and authorship limits are not necessarily going to help with the problem because the problem is too many people submitting a few papers, rather than the other way around.
Deepak Subramani retweeted
It's not just academia. It's young people (still in school/university (but not affiliated with professors/research groups)) that are trying their luck at getting something through as ICLR/NeurIPS/ICML papers are credentials for jobs these days. You can't even get a junior AI role without it.