Devansh

product @youtube & @GoogleDeepMind building YouTube's algorithm, AI products | before: news/discover, search ranking, ads | cs & econ @yale

[Devansh @devanshtandon_]

so cool!

i spend O(100m) tokens daily working with agents through interactive explainers

core loop:
1/ search: code / papers / documentation
2/ visualize html to understand
3/ kick off coding agents to build
4/ understand agent progress through dashboards

[Jeff Dean @JeffDean]

2/
Check out how Gemini 3.5 Flash instantly digests dense academic papers and autonomously codes a fully interactive, visual website explaining the intricacies of the research. It's an incredible stress test that seamlessly merges massive long context, deep reasoning, complex coding, and ultra-low latency.

It really helps you distill papers down to their essence and aid your understanding!

[Devansh @devanshtandon_]

@michelletomkim @techreview funniest moment of the day

[Devansh @devanshtandon_]

@MikeIsaac funniest moment of the day

[Devansh @devanshtandon_]

@madiator CC search tool use could be so much better, big gap vs gemini/others.

[Devansh @devanshtandon_]

@tanayj still so early — recs models have the same scaling laws as LLMs, LLM based recommenders are coming soon

[Devansh @devanshtandon_]

@zan2434 @daniel_levine @eddiejiao_obj @drewocarr super cool! congratulations

[Devansh @devanshtandon_]

Jensen highlights China's dominance of the open source frontier, the only AI CEO who can speak the truth

China disagreements hinge on:
1/ takeoff timelines
2/ taiwan timelines

Biggest mystery: what was Mythos really trained on?

[Dwarkesh Patel @dwarkesh_sp]

The Jensen Huang episode.

0:00:00 – Is Nvidia’s biggest moat its grip on scarce supply chains?
0:16:25 – Will TPUs break Nvidia’s hold on AI compute?
0:41:06 – Why doesn’t Nvidia become a hyperscaler?
0:57:36 – Should we be selling AI chips to China?
1:35:06 – Why doesn’t Nvidia make multiple different chip architectures?

Look up Dwarkesh Podcast on YouTube, Apple Podcasts, Spotify, etc. Enjoy!

[Devansh @devanshtandon_]

just shipped adept on a random monday

[Claude @claudeai]

You can now enable Claude to use your computer to complete tasks.

It opens your apps, navigates your browser, fills in spreadsheets—anything you'd do sitting at your desk.

Research preview in Claude Cowork and Claude Code, macOS only.

[Devansh @devanshtandon_]

If including AI-powered products, consumer feeds like Instagram, YouTube, TikTok should be at the top of the list.

Just like ChatGPT/Gemini run a LLM inference to answer user prompts, the IG/YT/TT feeds run a LLM / LLM-inspired inference to recommend content for users to consume.

[Devansh @devanshtandon_]

@_clarktang true! the depreciation schedule / tech obsolescence of older GPU generations is also overstated. older GPUs/TPUs will continue to run hot at 80%+ utilization, with great ROI, until the chips break down

[Devansh @devanshtandon_]

@milesgrimshaw I'm using ~$100/day most days on claude code

[Devansh @devanshtandon_]

wild to remember that vibe coding was coined just a year ago, now claude code writes most of my code

[Andrej Karpathy @karpathy]

There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard. I ask for the dumbest things like "decrease the padding on the sidebar by half" because I'm too lazy to find it. I "Accept All" always, I don't read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I'd have to really read through it for a while. Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away. It's not too bad for throwaway weekend projects, but still quite amusing. I'm building a project or webapp, but it's not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.

[Devansh @devanshtandon_]

Personal context is the new moat in consumer AI. Gemini’s advantage is that it can pull from Google Photos, YouTube, Gmail, Calendar. Imagine travel recommendations from an AI that knows where you’ve been (Photos) and what destinations you’ve been watching travel vlogs about (YouTube).

Gemini personal intelligence is how they catch ChatGPT – not with a better model, but with better context.

https://t.co/nxFBjBGyHd

[Devansh @devanshtandon_]

LLMs operate over O(1m) tokens of context

User input is O(1k): they won't type/talk beyond that
System prompts are O(10 – 100k): traditional prompt engineering territory
Context fills the rest ~900k: managing context effectively is critical & high leverage

[Devansh @devanshtandon_]

@deedydas manus has the best AI capability vs. consumer reach ratio (only second to claude imo), a great acquisition for meta

[Devansh @devanshtandon_]

great acquisition! manus is a powerful AI app, but so constrained by distribution. highest capability to consumer reach ratio, great for meta.

manus + personalization with context from meta's data is the foundation for a killer consumer AI app.

depending on price, could be the best hyperscaler acquisition of '25
(character → google was best of '24)

[Manus @ManusAI]

Manus is entering the next chapter: we’re joining forces with Meta to take general agents to the next level.

Full story on our blog: https://t.co/huPrnbITCi

[Devansh @devanshtandon_]

@grutzik_petra crushing it!!

[Devansh @devanshtandon_]

@tanayj grok: brave

(it’s the cheapest)

[Devansh @devanshtandon_]

for details: https://t.co/036PS2jqmc

[Devansh @devanshtandon_]

X is moving to a LLM-based recommendation system, built on Grok.

Feeds are the underhyped consumer application of LLMs, larger than search & chatbots. Across YouTube, Instagram, TikTok, X, the content feeds will be rebuilt on LLMs.

My talk at AI Engineer predicted this transition & shared research for YouTube's recommendation system rebuilt on Gemini.

[Elon Musk @elonmusk]

The 𝕏 recommendation system is evolving very rapidly. We are aiming for deletion of all heuristics within 4 to 6 weeks.

Grok will literally read every post and watch every video (100M+ per day) to match users with content they’re most likely to find interesting.

This should address the new user or small account problem, where you post something great, but nobody sees it.

We will also be adding the ability for you to adjust your feed temporarily or permanently just by asking Grok.

[Devansh @devanshtandon_]

@elonmusk When will Grok start creating personalized content on X?

LLM-based recommendation systems are the largest consumer application of AI – using grok to curate the X feed. Expect experiments with creation soon.

ref: https://t.co/036PS2jqmc for YouTube / GDM work here