4x Founder & CEO at Rocket š | Building the future of recruiting | AI hopeful
San Francisco, CA
Joined January 2008
- Tweets633
- Following1.1K
- Followers666
- Likes21.8K
Arjun Lall retweeted
Fewer tokens, same work. Particle-1 from @particle_api uses 28ā52% fewer output tokens than DeepSeek V4.1 Flash across four coding benchmarks, per Particle. Now on OpenRouter.
Particle-1 from @particle_api is live on OpenRouter.
A token-efficient coding model post-trained on DeepSeek V4.1 Flash. Particle reports it uses ~45% fewer tokens on average with accuracy close to the base model, so agent tasks finish sooner and cost less.
$0.25/M input, $1/M output, 1.05M context.
openrouter.ai/particle-ai/paā¦
Arjun Lall retweeted
Today is full of releases of insane optimism and optimization.
Suhail has the track record to build fantastic products.
I believe 2027 will be the year we keep twisting our heads basically daily.
weāre just getting warmed up.
1/ Excited to announce our new company Particle
This is our first step in a series of models focused on developers.
With particle-1, we focused on token efficiency and managed to keep parity in performance while using ~45% fewer tokens.
Live now: particle.ai
Arjun Lall retweeted
Particle will be on the frontier for smartest model at lowest dollar per unit of intelligence
1/ Excited to announce our new company Particle
This is our first step in a series of models focused on developers.
With particle-1, we focused on token efficiency and managed to keep parity in performance while using ~45% fewer tokens.
Live now: particle.ai
Arjun Lall retweeted
Agent swarms are a big unlock.
One agent can only do so much.
Many together woven together by artifactory-hack-style message board can do a lot more.
Arjun Lall retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Arjun Lall retweeted
Grep just achieved SOTA on the three major deep research benchmarks, beating Perplexity, Google, Nvidia, OpenAI, and Anthropic.
We're a two-person founding team.
Arjun Lall retweeted
Since whole brain emulation is looking a bit less remote this morning, this is a great time to read @qntmās incredible flash fiction MMAcevedo. 5 minutes of your time, but will haunt your nightmares for years.
qntm.org/mmacevedo
šØNobody wants to hear this but it needs to be said.
> Scientists just copied a fruit fly's brain into a computer. Neuron by neuron. No training data. No machine learning.
> It woke up and started walking. No one taught it to walk. No one trained it. No gradient descent. It just... knew what to do.
A fruit fly brain has 140,000 neurons.
A human brain is around 86,000,000,000.
And we've gotten really good at scaling.
Meaning with this proof, the first digital human won't be built by OpenAI. It'll be copied from someone who's already alive.
Your consciousness is software. And someone just proved it can be copy-pasted.
Start your day with that.
Arjun Lall retweeted
I think agentic AI would work much better if people took lessons from organizational theory, which has actually spent a lot of time understanding how to deal with complex hierarchies, information limits, and spans of control.
Right now most agentic AI systems seem to pretend that models have basically unlimited ability to manage subagents when that is clearly not true. We need measures of spans of control for AI. A human tops out at less than 10 direct reports. I am pretty sure that 100 subagents is too much for an orchestrator agent - suspect we need middle management agents (yes, I get it, insert middle management joke here).
Similarly, we need more attention to boundary objects. These are what is handed between groups (marketing to IT to sales) in organizations to convey meaning as a project crosses group boundaries, like a prototype or a user story. Right now agents pass raw text & maybe code back and forth. Structured boundary objects that multiple agents of different ability levels can read and write to would solve a huge number of coordination failures & reduce token use.
I also think aboht coupling, which is how tightly units inside organizations are bound. Most agentic systems are either too tightly coupled (every step needs approval) or too loose (Moltbook). This tradeoff is well-studied in organizations, I bet a lot would apply to agents. Other known issues like bounded rationality also apply, I suspect.
Everyone is rushing towards the (terribly named) agent swarm, but the issue wonāt just be how good the model is, it will be org design choices. I am not sure the labs see this, but we definitely need a lot more experiments with organizing agents done by people who understand real coordination issues.
Arjun Lall retweeted
š£ Open call to agent builders: Let's read agent skills from `.agents/skills`, so people don't have to manage separate folders per agent.
Today we pulled the trigger for Codex to read `.agents/skills`. Goal is to deprecate `.codex/skills`.
Pls like/tag/RT for momentum.
Arjun Lall retweeted
I've been using Claude Code for a few months now and I keep thinking about why it feels so different from everything else.
Everyone's trying to build "AI-powered" versions of existing apps. AI Photoshop. AI spreadsheets. AI whatever. More interfaces, more buttons, more layers between you and what you're actually trying to do.
Boris went the opposite direction. He put Claude in the terminal.
Most people think the terminal is this scary technical thing. And yeah, it used to be - you had to memorize commands, remember syntax, know exactly what to type.
But I realized: the terminal is just text input and text output. It's basically a chat interface. It's been sitting there the whole time, connected to everything on your computer - your files, your system, your processes.
Everything.
The problem was never the terminal. The problem was that computers only spoke computer language.
Now they speak English.
You can ask Claude Code to find a file. Clean up your desktop. Check what's using memory. Write some code. Whatever. It understands what you mean, translates it to what the computer needs, and shows you what happened.
No new app. No new operating system. Just your computer, but you can finally have a conversation with it.
I keep seeing people try to build the "AI OS" and honestly, I think they're missing the point. We don't need a new operating system. The one we have is fine. We just needed it to understand us.
That's what Boris figured out. The CLI was always the perfect interface for this. Direct access to everything. Just needed to speak human.
Watch his videos if you haven't. The clarity of thinking is rare.
Arjun Lall retweeted
There'll always be more emails in need of reply, more meetings to attend, and more updates to read. A person can fill the entire workweek with these tasks over and over again. But to stay sane and sharp, you must pay yourself first by doing the work that actually means something to you.
I feel this acutely as someone responsible to employees, customers, followers, and readers. I could do nothing all day but check up on projects, people, and posts, but my brain would quickly check out if it was just doing that.
So quite frequently, I just don't. Don't check in, don't check up, and instead dive into the work that checks my own intellectual boxes. Programming for the love of it. Experimenting for the hell of it. Researching for the fun of it.
In another age, I might have been tempted to apologize for such privilege, but screw that. Privilege is wonderful. You should do your best to earn more of it. Even if you have to carve it out of the bare rocks around you.
Ironically, the best way to do that is also to choose to always pay yourself first, however little at first. By solving your own problems, tickling your own interests, chasing your own curiosity. That's where you'll find the motivation to elevate your talent. To turn interest into competency.
And once you've developed some competency, you'll be rewarded with more privilege to build it further. This is the virtuous circle of merit.There'll always be an endless list of work that could be done.
You'll never get through it all and onto your own priorities, if you continue to put them at the bottom.
Arjun Lall retweeted
I love when Claude tells me something will take 6 weeks.
No dude, we are doing it this afternoon.
Arjun Lall retweeted
Anthropic co-founder Ben Mann on why he chose Montessori for his daughter:
"If this were 10-20 years ago, I'd be lining her up for top-tier schools and extracurriculars. But now I don't think any of it's going to matter. Learning facts is going to fade into the background. What matters is that she's happy, thoughtful, curious, and kind."
Ben Mann (@8enmann) left @OpenAI in 2020 with the entire safety team to co-found @AnthropicAI (now reportedly valued at over $100B).
In a rare interview, Ben opens up about:
šø What he saw at OpenAI that convinced him to leave
šø The AI nightmare scenarios that concern him most
šø His take on Metaās $100M talent wars
šø Why he believes 20% unemployment is inevitable
šø His āeconomic Turing testā for knowing when weāve achieved AGIāand why itās likely coming by 2027-2028
šø How focusing on AI safety created Claudeās beloved personality
šø What three skills heās teaching his kids to thrive in an AI future
šø So much more
Listen now š
ā¢ YouTube: youtu.be/WWoyWNhx2XU
ā¢ Spotify: open.spotify.com/episode/2dhā¦
ā¢ Apple: podcasts.apple.com/us/podcasā¦
Thank you to our wonderful sponsors for supporting the podcast:
š Sauce.app āTurn customer pain into product revenue: sauce.app/lenny
š @Lucid_LinkāReal-time cloud storage for teams: lucidlink.com/lenny
š @Fin_aiāThe #1 AI agent for customer service: fin.ai/lenny
Arjun Lall retweeted
I donāt mean to be a broken record but AI development could stop at the o3/Gemini 2.5 level and we would have a decade of major changes across entire professions & industries (medicine, law, education, codingā¦) as we figure out how to actually use it.
AI disruption is baked in.
Arjun Lall retweeted
Iām starting to become convinced you actually become a better programmer by having the LLMs write the code, because programming is less about writing code and more about clear thinking, and free-form writing forces you to think. The code is downstream from that.
Arjun Lall retweeted
A balanced life has work, friends, family, health, hobbies, etc. An imbalanced life goes very deep in one at the expense of the rest.
I think you want life to have chapters of balance and imbalance. Happiness comes from balance, progress comes from imbalance.
Arjun Lall retweeted
More and more, I have lost conviction that āminimum viable productsā make sense for product development.
It makes no sense to release a product with the core flow and then dismiss its viability after the aggregate data says people arenāt using it.
Instead, founders should have a fundamental belief about what people wantāand they should keep iterating until that value is correctly surfaced.
In practice, this means you should keep a steady flow of new users as you add components to the product, seeing if it solves a core part of the activation and signup loop.
The only part that could possibly be āhalf-bakedā is those individual components. But even those need sufficient quality so there are no confounding factors that distort the signal.