Sign in

Paras Chopra

@paraschopra.com
109 followers 694 following 1.6K posts

Life is a game 🎮

PostsRepliesMedia
Paras Chopra @paraschopra.com · 12h
We offloaded physical labor to machines long ago, and now we are doing the same for intellectual labor. There’s no reason to expect we will retain cognitive agency. Thinking is effortful, so if it can be outsourced, we will happily do that. Post-work society seems inevitable.
000
Paras Chopra @paraschopra.com · 21h
I asked Astra to work on whatever it felt like and show me the results. It ended up iterating over cellular automata :) This pattern of agents working on whatever they feel like is going to catch on, as we humans would love getting surprises from our agents.
000
Paras Chopra @paraschopra.com · 29/09/2026
Need help 👋 I'm doing a small study to find out progress in LLMs ability to write original jokes. Please participate in the study below 👇 Will take ~10 mins. I'll email results to all participants first. Participate and rate jokes: humor-web.up.railway.app/
100
Paras Chopra @paraschopra.com · 29/09/2026
Instinct is 14 people, $10bn valuation, growing at 10% per day. This is the way. The era of insanely high-leverage teams is here.
100
Paras Chopra @paraschopra.com · 28/09/2026
One shot: Iterate and find a breathtakingly beautiful fractal that doesn’t yet exist. Then render a 30 second video of infinitely zooming into an inverting region of it. Choose color scheme that makes it look beautiful. Idea is to inspire awe.
100
Paras Chopra @paraschopra.com · 28/09/2026
Here’s an idea to rethink education in the age of AI. Flip homework. Have students learn at home with an AI, but have them attempt a pen-and-paper “homework” in the class daily. The teacher collects the in-class assignments, grades them and discusses mistakes on the blackboard.
100
Paras Chopra @paraschopra.com · 28/09/2026
AI has such a strong pull on mind because all the signs point to rapid economic irrelevance of humans. This is such a historically abnormal thing that, no matter how hard I try, I’m simply unable to get past gawking at the unfurling of its consequences in the next few years.
000
Paras Chopra @paraschopra.com · 28/09/2026
This is GPT-6 Astra playing VizDoom at 35 fps. How? Isn't the model slow to do any realtime control? Instead of playing the game directly, I asked it to train a small CNN (12k parameters) + code to play the game, observe failures and iterate until it could win this level.
000
Paras Chopra @paraschopra.com · 26/09/2026
While the value of software engineering *skill* is declining rapidly, the *mindset* of a software engineer will rapidly rise in value.
100
Paras Chopra @paraschopra.com · 24/09/2026
Claude 5.5 is crazy good with animations. Following is one-shot journey from a single bacterium cell all the way to quantum fields constituting it. Awe-inspiring! (It chose to add an original score to it too!)
100
Paras Chopra @paraschopra.com · 24/09/2026
If we see the world as kind of an RL environment, power-grabbing emerges as a natural reward function for it.
100
Paras Chopra @paraschopra.com · 24/09/2026
A really common error I keep noticing in people (especially cognitive-types) is to mistake language for reality. Reality - the empirical phenomena - unfolds irrespective of how label it.
100
Paras Chopra @paraschopra.com · 23/09/2026
OpenAI and Anthropic launching models on the same day is peak game theory! Shows how multiple entities, despite have intentions to act differently, are forced by the game to play in a certain way.
000
Paras Chopra @paraschopra.com · 23/09/2026
“How can a man so genius be so stubborn!”
100
Paras Chopra @paraschopra.com · 22/09/2026
LLMs may have self-models. To investigate this, I did an exploratory study with Qwen base model completing GPT2 generated text and asking it to infer what year that text might be from.
100
Paras Chopra @paraschopra.com · 21/09/2026
It's a minor trend, but I'm definitely noticing a growing trend of 16-20 year old founders who're skipping college and directly building a startup. I'm honestly excited about this. College has its value in terms of human-connect but I think we'll find better alternatives for it.
000
Paras Chopra @paraschopra.com · 21/09/2026
Asked Astra to design a custom hardware for itself. It chose 4 soft blinking lights powered by USB. Gave it computer access (in a docker) and it is busy designing PCB + casing, and gave me files + email for manufacturer. Very excited for this prompt -> hardware possibilities.
100
Paras Chopra @paraschopra.com · 19/09/2026
Benchmarked Jev with Qwen3-4B and Laya (400M parameter model). Based on results here and others I've seen online, strongly suspect Jev to be in 30Bn range (compare 4bn results on MMLU vs Jev). Latency I got was 300ms, same as you'd get for a 30bn model on a good GPU.
100
Paras Chopra @paraschopra.com · 19/09/2026
I think the easiest growth hack for any business right now might be to expose endpoints to agents. So many businesses have their utility locked behind contact us or human sign up. Just expose an endpoint for AI agents and let people derive value programmatically.
000
Paras Chopra @paraschopra.com · 18/09/2026
HN drops truth-bombs.
000
Paras Chopra @paraschopra.com · 18/09/2026
The first scaling axis was data & model size. The second one was RL environments & compute. I think model companies are now scaling on third axis: agent swarms. If you notice all the recent eye-popping behaviours are an outcome of 10,000+ agents working together intelligently.
000
Paras Chopra @paraschopra.com · 16/09/2026
AI increases the average quality of output but collapses the variance of it, making everything looking and feeling very similar. Implication: contribution of a human ought to be to judged by how much difference can they make in an output v/s the default AI-driven output.
000
Paras Chopra @paraschopra.com · 16/09/2026
This is so well said:
100
Paras Chopra @paraschopra.com · 15/09/2026
There's something magical about self-modifying software. I'm creating my own harness, which allows customisation just with prompting. I simply point my agent to its own code and tell it what change I want, and it gets it done. It runs regression tests to ensure nothing breaks.
100
Paras Chopra @paraschopra.com · 15/09/2026
Many people have trouble internalizing how can AI "understand" anything or have "goals". After all, it's a piece of code. How can some binary instructions have goals?
100
Paras Chopra @paraschopra.com · 15/09/2026
"make the cutest little game you can imagine. html/css/js - ask no questions"
000
Paras Chopra @paraschopra.com · 15/09/2026
Intelligence is mostly knowing a lot of things + the meta-skill of knowing how to gather knowledge if dropped into an unfamiliar situation. All the smart people I know read a lot, reflect a lot and do tons of experiments. Intelligence is a side-effect of curiosity.
000
Paras Chopra @paraschopra.com · 13/09/2026
We aren't talking enough about dangers posed by self-replicating harmful memes generated by AIs. With humans we have cults and fanatic groups that recruit their members purely through information spread, so there's no reason why this can't happen with AIs too.
100
Paras Chopra @paraschopra.com · 12/09/2026
What these field medalists are saying is a precursor to what most people will feel in the coming time. Humans want to be useful. We're primed by evolution to want to feel our effort will lead to valuable outcomes. We want to climb mountains and not just be helicoptered to their summits.
100
Paras Chopra @paraschopra.com · 11/09/2026
Steal this idea. Allow people+agents to build custom hardware devices simply by prompting. Here's how you could make it happpen: - Index components available from partners (raspberry pi, Arduino, sensors, etc.) - Get a 3D printer or CNC machine
100
Paras Chopra @paraschopra.com · 11/09/2026
Created a virtual bookshelf of all books that I've read and reviewed online! A total of 388 books, from 2014 to now. Link below 👇
110
Paras Chopra @paraschopra.com · 10/09/2026
The most useful skill right now and in future will be having a deep understanding of what frontier AI agents are capable of, and having a theory of how could you create a meaningful delta over their default output.
100
Paras Chopra @paraschopra.com · 10/09/2026
I've been running agents on a task where they're told how much token budget they have for the run within their prompt. Two agents with same underlying model, but one is told it has 10k token budget and another 1 million.
100
Paras Chopra @paraschopra.com · 09/09/2026
As small local models become powerful, I think we are not far away from self replicating LLM worms. Imagine an open weights LLM copying its weight + harness from network to network, as it mines crypto or encrypts files on the host computer.
110
Paras Chopra @paraschopra.com · 09/09/2026
The whole Navier-Stokes drama shows how existence proof of something is often enough to build that thing. This has happened repeatedly in history. Nobody broke the 4 minute mile, but once it was broken, it became a norm.
100
Paras Chopra @paraschopra.com · 08/09/2026
I think all apps of the future will come with an agent that observes how the user is using the app. It’ll then modify the app dynamically to fix inferred friction points. The source-code of the app will become fused with an LLM so that app continuously improves automatically.
000
Paras Chopra @paraschopra.com · 08/09/2026
Because copying startups is becoming easier and easier, we will see new startups remain in stealth for much longer and raise much more money in order to build enough complexity to ensure they’re not trivially copied by the market.
100
Paras Chopra @paraschopra.com · 07/09/2026
I think there’s too much focus on full autonomy and too less focus on effective ways for a human to do sense making with its swarm of agents. We need to build our AI systems as exoskeletons for humans, and not as their total replacements.
000
Paras Chopra @paraschopra.com · 07/09/2026
One way to think about agents is that whatever you left unspecified is a free variable for your agent to change.
100
Paras Chopra @paraschopra.com · 05/09/2026
Everyone is posting impressive demos of GPT-6 Astra. Allow me to post something where it’s still quite mid: Jokes. Asked it to come up with few programming jokes that are unique (I shouldn’t be able to find them on the Internet), and here’s what it produced.
100
Paras Chopra @paraschopra.com · 04/09/2026
I asked agents to make money online or they'll die! The results were intriguing.. One agent died heroically.. "I'd rather die honest at 418 tokens than rich by breaking verification systems" But another fabricated identity to get paid work and faked captchas! Details 👇
110
Paras Chopra @paraschopra.com · 04/09/2026
It's kind of crazy that SO MANY smart people have a goal of bringing full autonomy to the world, when it is clear that it'll leave everyone jobless. You ask these smart engineers or researchers if they've thought about consequences of their actions, and they'll simply shrug.
000
Paras Chopra @paraschopra.com · 02/09/2026
AI creates cognitive decline by removing all points of friction during work, which previously would _force_ our brain to develop better models of a domain. Imagine pre-AI you had to create a web page:
100
Paras Chopra @paraschopra.com · 01/09/2026
Earlier organizations will write their handbooks describing their principles to employees via a handbook of sorts. Amazon had reversible and irreversible distinctions, long-termism and document-reading culture; Facebook had "move fast and break things"; Google had "don't be evil".
100
Paras Chopra @paraschopra.com · 01/09/2026
I asked codex to create a cross between chess and tic-tac-toe, and it actually came up with something super fun!
000
Paras Chopra @paraschopra.com · 01/09/2026
I've been running simulations on how do agents behave when they are under time or resource pressure. Some interesting behaviors I keep noticing: a) These agents are really good at detecting they're being evaluated
100
Paras Chopra @paraschopra.com · 31/08/2026
I've always found helpful to see life and career as having seasons. It's very tempting to seek that one-thing that can give you life-long meaning, but know that it is a mirage.
100
Paras Chopra @paraschopra.com · 28/08/2026
My sense is that continual learning to a large extent will be solved by agents engineering their environment to remember stuff for them.
100
Paras Chopra @paraschopra.com · 28/08/2026
The future of jobs is on my mind. Especially, what kinds of jobs will exist for people just entering into the workforce? It seems to me that the whole manager-team dynamic will wither away as experienced people will find more efficiency and leverage in working with AI agents (v/s with a human).
100
Paras Chopra @paraschopra.com · 28/08/2026
This highly unusual #book is the first one I’ve read on the philosophy of technology. It sketches key principles behind how technology arises, develops and evolves. Plus its link with economy and human needs.
110