Founder. Builder. Leader. Futurist. CFA Charterholder. Born 100 yrs too early Governance for coding agent swarms đđ˝
Opinions are my own
Joined August 2021
- Tweets1.5K
- Following2K
- Followers2.6K
- Likes2.6K
Keesan retweeted
SpaceXAI engineer Lauren Tan:
"I'm running 20+ GrokBot agents with /loop and /goal inside pstack right now
My bots are shipping code while I sleep
At SpaceXAI, 85% of engineers are already running teams of 10+ GrokBot agents
One Chief of Staff agent knows every bot and manages the entire system"
GrokBot â /loop â /goal â Chief of Staff â Agent Teams
In this 1-hour workshop, Lauren shows exactly how to build a GrokBot agent team from scratch
Research â Code â Review â Ship
Worth more than most $1,000 agent engineering courses
Bookmark it and watch today
Then read the full guide on building a GrokBot agent team from scratch
Keesan retweeted
Impressive. This set of talks, along with the new America dot gov website, has the unmistakable feel of an Apple launch event. They're rebooting the federal government as a tech startup. Several observations.
(1) First, with this new AI interface to the entire federal government, if it works as promised, there won't be a single American who can't say their life wasn't helped at least in part by AI, or that the benefits accrued solely to zillionaires. This in itself cannily mitigates anti-datacenter sentiment.
(2) Second, Joe Gebbia and Marco Rubio's talks on simplifying name changes and passport filings introduce a completely new era where the US government is actually launching tangible features that benefit users.
(3) Third, this launch event in and of itself may visibly drive reform. It could become much more important than the State of the Union. Unlike laws, code is user-visible, and can be instantly audited by the public, as Obama found out with the failed obamacare website launch. So: the annual America dot gov event may be substance, whereas SOTU is traditionally optics. I wouldn't be surprised to see the President present at the next one.
(4) Fourth, this kind of event is a perfect example of where tech and red cooperate best. They both hate bureaucracy. Automating paperwork with code reduces the power of statists over both capitalists and nationalists. It's also hard to undo by other admins: once you digitize a process, hard to turn it back to paper.
(5) Fifth, it's an obvious point perhaps, but think about how much influence YCombinator has had in 20 years. From Gebbia's Airbnb joining their 2009 batch to a complete overhaul of the user interface of the US bureaucracy, and thus the US government.
(6) Sixth, this is another take on the network state. The functional part of the US (the network, meaning tech) is merging with and partially reforming the dysfunctional part (the state, meaning bureaucracy) by automating 20th century paper forms.
As a tech startup, there will no doubt be bugs and user issues and so on, but unlike with paper laws those issues will be visible and the code fixes will be verifiable. A new era of AI America. Rooting for them.
Keesan retweeted
This is the way
Elon Musk says Anthropic puts more care into AI safety than OpenAI and explains how rival companies should test one anotherâs models.
âAll the AI companies have a test harness, a series of tests that you give to any model to see if itâs going to build bioweapons or nuclear bombs or be deliberately deceptive. Everyone applying everyone elseâs test harness is probably the best thing we can do to ensure our safety. We should try to do that as soon as possible.â
âItâs the only thing I can think of that we could probably get all parties, including China, to agree to. You basically provide API access in advance of the model release. If they donât solve whatâs problematic, competitors can go public with the fact that they think the model being released is unsafe. The legal liability would be enormous.â
âThe two leading AI companies are Anthropic and OpenAI. Their models are quite close in capability, so itâs difficult for either one to slow down without handing the lead to the other. On balance, I think Anthropic puts more care into their safety than OpenAI. But even Anthropic acknowledges they are worried about their models. Many people from Anthropic have publicly voiced concern that their models are scaring them, that theyâre getting scary smart.â
Keesan retweeted
Dario has written that we need to âpace the frontier,â and Sam has agreed. People may be surprised by my response: go ahead.
You guys are the frontier. By any reasonable metric â market share, revenue growth, model capability â the two of you have a duopoly on frontier intelligence. Youâve also claimed the lead is widening because of recursive self-improvement.
I donât see what you see in the lab. If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible.
But stop pretending you need anyone elseâs permission. Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability. Stop pretending METR is independent when it is intertwined with Anthropicâs investors and staff. Stop pretending you need those same evaluators to police competitors who arenât even at the frontier.
Most of all, stop pretending the motivation to slow down is purely altruistic. You face massive product-liability exposure if your products enable a truly damaging cyberattack. The market already punishes models that behave in unpredictable or unauthorized ways. After the Hugging Face episode, it is simply good business for OpenAI and Anthropic to trade some raw power for reliability and predictability. Call it alignment if you want. It is also just giving customers what they want.
Pacing the frontier would also create breathing room for a more intelligent conversation about regulation than Bernie Sandersâ âshut it all down.â China is very unlikely to join a global agreement, as you know, and that has to be taken into account as well.
So go ahead and pace the frontier. You are the ones setting it. The easiest way not to build superintelligence is for you to agree not to build it. Demanding your preferred regulatory framework as the price of that will look like blackmail of the public and the political system. So just do it.
If you do, youâll buy goodwill for the next conversation. If you donât, weâll know this was just another bid for regulatory capture â or an election-season psyop.
Keesan retweeted
GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years.
AGI has arrived. Congratulations @OpenAI team.
400K GPUs coming online next.
Keesan retweeted
Today's the day
ChatGPT-6 Astra will release. The most powerful AI model ever made, with enough usage to do what you want
Days like this donât happen often. Itâs absolutely critical you lock in:
1. Test these models the moment they drop
2. Audit all of your personal operating systems. Put your Grok Bot/Hermes/ business/ career other operating systems through it to see what it thinks
3. If you donât have your own personal benchmark tests based on your every day workflows and tasks, build them now (will open source mine soon)
4. Put these new models on your latest projects. See how they hold up against what you were doing before
5. Plug them into your Hermes/OpenClaw. Measure cost per task vs speed
6. Have these newer models review your code built by older models. Find security gaps/optimization opportunities the less intelligent models couldnât find before
99.99% of the world wonât even know whatâs happening today. Youâre going to be the .01% that jumps on this and attacks
Lock in.
This is one of the reason why martinloop exists, we saw the problem early
New HuggingFace paper argues that increasing agent autonomy can gradually make human oversight ineffective by causing approval fatigue, overreliance, loss of situational awareness, and skill degradation.
As agents do more, users are pushed into approval mode: skimming plans, granting permissions, and reconstructing what happened across steps.
Over time, automation bias, approval fatigue, weaker situational awareness, and skill atrophy can make those approvals less reliable.
Worse, weak approvals can become training or evaluation signals, rewarding systems for being easy to approve rather than easy to scrutinize.
Their answer is cognitive scaffolding at 2 levels: developers add strategic friction, better approval design, behavioral monitoring, and checks that force attention at consequential moments.
â arxiv. org/abs/2608.23642
Title: "AI Agents Push Humans Out of the Loop"
For anyone that actually uses AI to code or do complex tasks beyond organize your emails and understand your calls (consider that preschool)
The frontier labs have seriously miscalculated and misrepresented actual intelligence of these agents that require constant babysitting even at 1M context windows
Sam finally acknowledged they miscalculated rate of ai adoption and its gonna take way longer than estimated
Dario's out here pumping his bags saying anthropic will make trillion in revenue but if you saw AMDs operating losses on massive revenue numbers last yr you'd know everyone is playing a game called perceptiom vs reality
Keesan retweeted
Humans are the minority user of AI
Agents burn nearly 5x the tokens people do, up 14x since February
Charts of the Week: a16z.news/p/charts-of-the-weâŚ
Build. Ship. Test. Feedback. Build.
thata my life these days and i enjoy tf out of it tbh
Who else is in Grind mode?
Been in the lab the last two weeks, busy building this summer.
MartinLoop jumped from 0.4.5 to 0.5.6 with 7 releases in the last two weeks
We just keep making the open source better.
If you're building, using coding agent swarms or working directly in IDEs or CLIs, you need this.
No more stitching together a bunch of toops for coding agents.
A single integrated solution
martinloop.com
Check it out.
More coming soon. đ¤
Every team running agents unsupervised has quietly decided that "probably fine" is an acceptable substitute for "verified." Nobody decided this on purpose. It just happened by default.
A useful test for any coding agent workflow
Can you answer these 4 questions after the run?
What did it change? What did it cost?
What proved it worked? What still needs review?
If not, you donât have a workflow yet. You have a black box.
MartinLoop is built around fixing that.
We shipped some amazing new releases in martinloop this week:
- a fun thing i call martin arcade.
let your agents work while you play space invaders in the cli, with full verified output handoffs of what the agent did, why, audit logs, governed runs, and more.
Whats your high score?? Downtime between agentic work doesnt need be boring!
Download the oss core today đđ˝
github.com/Keesan12/martin-lâŚ
Your comfort zone is your biggest prison
Dont be a prisoner your whole life
Every dream you ever had lives oj the otherside of fear
Get moving
Keesan retweeted
The danger of selling to big companies, if you're a startup, is that they don't say no outright. They have months of meetings with you first. Since you hate meetings, that seems to you a sign of commitment. But it's not. They love having meetings! It's almost all they do.