SWE-III @Google | Ex- @paloaltontwks | Web3 | Anime | Gaming | Chess | Movies | Series | SciFi | Cricket | Cars | Coding | A pinch of randomness. Views my own.
Bengaluru, India
Joined April 2017
- Tweets2K
- Following386
- Followers1.4K
- Likes6.7K
Venkatesh Dhongadi retweeted
a lot of people fail to realize that anyone with a bunch of followers on social media is still just a normal fugging dude
being mean and hostile to another human even if they have a big following on social media doesn't change they are just ONE DUDE(or dudet), just like you.
Sometimes I forget that hundreds of thousands of people are following what I do.
A few years ago, I could make a decision, share a random thought, or say something stupid, and nobody would notice.
Now everything has a much bigger impact.
And I don’t think my brain has fully adapted to that yet.
I’ve always tried to be as raw and authentic as possible on here. I share what I’m building, what I’m thinking, my wins, my failures, sometimes without thinking too much about how people will perceive it.
That’s what I’ve always done, and I don’t want to lose that.
But I’m also just a human who sometimes gets things wrong, even with good intentions.
It’s a weird feeling trying to stay the same person while realizing that the things you say and do carry a lot more weight than they used to.
But I can't complain! I'm incredibly grateful to be where I am today, and I wouldn't trade it for anything.
I guess I just have to learn how to navigate all of this while staying true to myself 🫂
A 100K context window beats 1M tokens for long-running agents.
Every team building autonomous coding agents hits the same wall around hour 6 of continuous session:
Stuffing 1M+ raw tokens into eveyr turn degrades multi-hop retrieval ("lost in the middle"), spikes TTFT latency, and burns API budgets on stale tool outputs.
Here is how the trending billion-context architecture keeps month-long sessions sharp inside a strict 100K active window while cutting token spend by 5x:
1. Tiered Episodic Eviction (Raw -> Digest -> Anchor)
Instead of appending every bash log, stack trace, & file read verbatim, old turns are compacted into structured state deltas. Raw stdout disappears once the diff or test assertion passes.
2. Working Set Pinning (<= 100K Active Budget)
Treat the LLM context window like CPU L1/L2 cache, not a hard drive:
- L1 (Pinned): System invariants, active task graph, & Current file ASTs.
- L2 (Rolling): Compressed summaries of completed subtasks & failed hypotheses.
-L3 (Cold store): Full historical transcript indexed on disk for on-demand pointer lookup.
3. 5x Lower Token Cost Across Billion-Token Lifecycles
When an agent runs across weeks and billions of cumumlative okens, prefill cost dominates total spend. Compressing completed steps at task boundaries keeps attention heads focused on live code constraints instead of 400 lines of yesterday's webpack warnings.
Stop paying to re-read dead terminal logs on every prompt turn. Desing your agent memory like a virtual memory pager.
Bookmark this for your next architecture review.
2
5
272
Open-source repo & architecture notes for billion context : github.com/ranxianglei/billi…
Key takeaway if you run long-horizon Claude Code or custom agent loops: compacting context at subtask boundaries gives you higher reasoning accuracy at 100K tokens than dumping raw unpruned logs into a 1M window.
1
47
Hi @zomato
I can’t get through your shitty bot for help on the app, perhaps I’m posting here. I’d like a refund for the way my pizza is delivered. Spoilt.
Do better. Order : 8831168820
@zomatocare
Folks, matter solve hogya hai (in < 24 hrs).
Thank you team @zomato @zomatocare 👏
Hi @zomato
I can’t get through your shitty bot for help on the app, perhaps I’m posting here. I’d like a refund for the way my pizza is delivered. Spoilt.
Do better. Order : 8831168820
@zomatocare
This has got to be one of the funniest edit on the internet! 🤣 This is one of my fav songs
How are people able to even get this accuracy?
Source: IG.
1
6
446
Gemini 4 Carbon has been at it since 3 hrs, has provided me with 23 updates & I'm afraid I'm gonna hit my weekly limit by the time this is done.
If the limits get exhausted, I'll switch over from Antigravity to Claude Code & make Fable 5.2 pick up where we left off.
I'm building a game which I'm gonna absolutely love playing & I hope the rest of you will enjoy playing it too xD
About damn time!
I think I’m just gonna build this game myself. I’ll make a video on how I do it.
The night’s young & I’m only getting started :)
10
4
516
When I said 23 updates, what I meant to say is, after making some progress it's dropping a short paragraph of work done & is proceeding to work on the next task.
This is a one shot prompt btw. And rn, we're are at update #30
1
1
128
I absolutely love /swarm.
Having used it internally, given a proper plan, it does a phenomenal job with writing the code, reviewing code (multiple rounds) and testing all while following the code guidelines, making my PRs squeaky clean and ready to merge with me rarely getting a code review feedback on my PRs.
Gemini 4 Carbon is now available for Ultra users in Antigravity (and yes, we will quickly expand access to more users). We are also launching /swarm as a native experience to leverage multi agents to complete complex tasks.
Swarms have been a step function improvement for capability internally, especially with Carbon.
1
5
452
About damn time!
I think I’m just gonna build this game myself. I’ll make a video on how I do it.
The night’s young & I’m only getting started :)
A dev just lost a final offer after blanking on production scenario questions.
AI writes syntax now, so teams test how you debug failing systems under load.
What is the toughest scenario question you have faced?
1
222
Vercel just open-sourced a browser CLI for AI agents.
48.8k stars!! Most playwright setups choke LLM context windows with raw DOM dumps. Agent-Browser gives Claude or Antigravity clean CLI primitives to test UI flows.
Bookmark this for your next architecture review. Added repo link in the comments.
3
9
383
GitHub repository for vercel-labs/agent-browser (48.8k stars today): github.com/vercel-labs/agent…
63
Nvidia CEO Jensen Huang just emailed all employees admitting an affair.
Now US Intelligence is probing whether the woman involved, an exec at a Singapore firm that bought $4.8B+ in NVIDIA chips, is linked to Chinese intelligence.
Earnings call is Nov 17.
Closed-door US Congress briefing is Nov 18.
Do you think this is just a personal matter, or a real risk to the AI industry.
6
328
The google search trend for meaning of the word "espionage" lately.
Who is gonna tell me? xD 🌚
1
2
291
