Memory Research @LittlebirdAI. Learning to think, and teaching it to machines, sometimes robots. Tweeting notes-to-self and musings.
New Delhi, India
Joined April 2020
- Tweets1.1K
- Following631
- Followers521
- Likes7.3K
Pinned Tweet
5 undergrads, no lab, no supervision, and 1 year of "figuring it out" later, we're hyped to present - "Chronocept".
It's the world's first benchmark to instill a sense of time in machines.
This our lives' best work, and we hope to say this about everything we do hereinafter.
Founder’s Office/CoS role if you have great taste in fashion + want to build a startup from the trenches.
I wish I could be more specific but it’s hard to reduce this role to a few “responsibilities”.
Love the energy at the office every time I drop in.
metafashion.in/careers/found…
I had such a fun chat with @Rifachhabra at @abplive. Here are a few cool bits from the interview! Thanks for having me.
By the way, we gave our website a nice overhaul. The brands section is my favourite part:
metafashion.in/brands
You can also sign up for our newsletter. The first one should be out soon, with cool updates on fashion on Roblox and a little about what we're building:
metafashion.in/#newsletter
We're also hiring a Founder's Office intern. Students from any university can apply. Honestly, this is the role I'm most excited to hire for:
metafashion.in/careers/found…
@Roblox
My latest paranoia is writing something without AI assistance, reading it back, and sweating “do I sound like an LLM?”
There are ~4400 engineering colleges in India. Presuming each college to have 20 technical clubs with an average of 40 members each, that’s 35,20,000 “enthusiasts” and 2,64,000 “core leaders”.
(This excludes non-technical clubs and non-engineering colleges with technical clubs)
Krish Goel retweeted
it's sad that some engineers will never have the pleasure of being absolutely ripped to shreds during a human code review
Love how krishgoel.com/listening looks right now.
Not sure about the underratedness of these but @goatgoel and I maintain this playlist: open.spotify.com/playlist/6H…
Cannot wait for the next logical step: a “derivatives market on other people’s life insurance”. Long/short humanity. God bless.
The value prop with Jev (@typesafeai) is it enables almost all pipelines and agent loops to be e2e differentiable (and trainable) at scale (with post-training) and that I think is really powerful.
Gave @cursor_ai Projects an honest shot with a multi PR undertaking for 3 days. Here's the review: great idea, crappy execution.
Abstracts away very important details, very hard to see the code, no clarity on associated artefacts (plans, PRs, diffs), no worktree management, no subagent statuses.
Got me to ship slop I eventually had to clean up manually. It's really addicting to keep sending messages and not worry too much - the end result looks good enough and the minimal communication from orchestrator makes it sound smarter than it is.
Someone should build a well thought out version of it.
The new Projects feature on @cursor_ai is super relevant to me (persisted memory + stacked PR management), but the UI is confusing:
1. Hard to tell where or why an agent is working in a given directory/worktree - the Agents/Projects window looks dir/branch agnostic
2. Subagent code changes don't show in the sidebar
3. The agent can't access its own files/plans, and hyperlinks to them often break
The new Projects feature on @cursor_ai is super relevant to me (persisted memory + stacked PR management), but the UI is confusing:
1. Hard to tell where or why an agent is working in a given directory/worktree - the Agents/Projects window looks dir/branch agnostic
2. Subagent code changes don't show in the sidebar
3. The agent can't access its own files/plans, and hyperlinks to them often break
Very hopeful about improvements and moving over my entire workflow to this which currently involves jumping between @conductor_build and @cursor_ai IDE
Krish Goel retweeted
There are a million meeting notes tools, and I'm obviously biased, but I think ours is the best.
Meetings need to be understood in the context of your life, there is a lot of information outside meetings that is relevant to the meeting, and Littlebird understands all that context and situates your meeting notes in the context of what was discussed on slack, email, imessage, etc.
You can also do things like tell Littlebird to complete as many of the follow-up tasks generated in the meeting as possible, and every day we can complete more of those tasks autonomously. For those of you that are already users, get ready for some mind blowing additions coming in the next few weeks :)
Littlebird notices when you're going into a meeting and gets ready to take notes.
Zoom, Google Meet, Slack huddles, whatever the call is on. It detects the meeting starting and offers to transcribe, so there's nothing to open and nothing to remember.
You can choose to turn off auto-start in Settings.
Krish Goel retweeted
“don’t write that down”
got it, adding to memory that i shouldn’t write that down
Krish Goel retweeted
Just like people ask "alignment to whom?", I feel like we need to ask, "recursive self improvement at what?"
Improvement is measured wrt some capability. Which capability? If the goal is to get unbounded amounts of that capability, we should be more mindful about what we choose
Krish Goel retweeted
Basically, no. If loops work, you’re probably writing low utility software that doesn’t do anything interesting or is producing voluminous technical debt
