What @pirroh and @amasad have been cooking here is super impressive and a really helpful harness to use swarms immediately.
Code is usually not the goal, it's the tool to support a business outcome, and systems of intelligence like replit /missions focus on the latter directly.
Agent swarms are the next scaling law in AI. We scaled pretraining, then reasoning. Now we scale the number of agents working together.
On 13 NumPy optimization tasks, 1 agent hit up to 4x speedups, 16 agents reached 17x, and 32 agents reached 19x.
Today we're launching Replit Missions to put swarms to work.
Wow! Look at that rainy day. Is that a cold I feel coming on? Probably best to clear my calendar and work from home today.
Good morning! Grand Theft Auto VI is one of the greatest games I have ever played. Review: bloomberg.com/news/features/…
Anish Acharya retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me.
I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA.
I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation.
Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.
Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours.
Jensen
Put together my list of the smartest thinkers on swarms right now, check them out here and let me know who i'm missing!
cosign.co/lists/illscience/s…
Swarms are quickly becoming one of the most important topics in Silicon Valley, sharing a few notes from recent conversations:
- The naive view of swarms is that they’re a sum-of-parts method of parallelization / brute force, but we’re empirically seeing the “shape” of swarm intelligence be very different.
- The way swarms achieve this is through a combination of parallelization, entropy, and specialization, leading to emergent solutions that aren’t covered by a small number of brilliant frontier models acting as central planners.
- It’s also not obvious that the economics of swarms follow higher cost = better output. Swarms that mix and match frontier and non-frontier models with different reasoning levels may be more economical than a small number of frontier agents, and broadly, there should be a new Pareto curve here.
- Swarms will be orders of magnitude more token-consumptive than agents, just as agents were compared to humans + models, further driving constraints in inference supply.
I think it’s important to build intuition here so have been running some experiments that I’ll share soon. A great way to get started with swarms is the /swarms command in the new google anti-gravity.
A few of the smartest thinkers on this topic are @alehandromz @jsnnsa @gokhan_egri @pirroh @amasad
Love to hear other POVs on swarms if folks have them!!
Tarun has quietly been building one of the most impressive products of this cycle. What a fantastic outcome for the team, I can't wait to see what you and @rauchg build together. đź’Şđź’Ş
4
94
4,901
Swarms are quickly becoming one of the most important topics in Silicon Valley, sharing a few notes from recent conversations:
- The naive view of swarms is that they’re a sum-of-parts method of parallelization / brute force, but we’re empirically seeing the “shape” of swarm intelligence be very different.
- The way swarms achieve this is through a combination of parallelization, entropy, and specialization, leading to emergent solutions that aren’t covered by a small number of brilliant frontier models acting as central planners.
- It’s also not obvious that the economics of swarms follow higher cost = better output. Swarms that mix and match frontier and non-frontier models with different reasoning levels may be more economical than a small number of frontier agents, and broadly, there should be a new Pareto curve here.
- Swarms will be orders of magnitude more token-consumptive than agents, just as agents were compared to humans + models, further driving constraints in inference supply.
I think it’s important to build intuition here so have been running some experiments that I’ll share soon. A great way to get started with swarms is the /swarms command in the new google anti-gravity.
A few of the smartest thinkers on this topic are @alehandromz @jsnnsa @gokhan_egri @pirroh @amasad
Love to hear other POVs on swarms if folks have them!!
Swarm riders and Swarm-native companies
AI-use even among highly AI-native people is power-law distributed.
There seems to be a new class of people — call them “swarm riders” — who are capable of consistently commanding large swarms of agents to productive ends.
This is a new kind of managerial skill, analogous to the skill it takes to manage large organizations of people, but wholly different in nature and of course more technical.
Based on the swarm riders we know (including many at @Tempo and @Paradigm), these people are capable of productive token spend that is 1000x-10000x higher than median software engineers, and 100x higher than top 5% Claude Code users.
The art of swarm riding is different from software engineering and more akin to machine learning, where practitioners grapple with black box systems: defining reward functions and evals, transmuting non-verifiable problems into verifiable problems, running experiments, etc.
Swarm-native companies are a new class of company founded by and populated by swarm riders. These companies are organized around swarms in a way that conventional companies will struggle to replicate, as the cost of internal human coordination is sand in the gears of an effective swarm. Effective swarm-native companies will define the right reward functions and evals, remove functional and bureaucratic barriers, and then let the swarms cook.
We’ve gotten to see a few swarm-native companies up close — reshaping financial rails @Tempo, building a new gaming platform @Spawn, cyber swarms at Cathedral Systems — and it’s clear these teams wield a significant advantage over pre-AI teams.
Some questions I’ve been wondering about:
1) How fast will swarm riding diffuse through the economy? The implications for token spend could be significant.
2) Can swarms be effectively adopted by existing companies or will it generally track the formation of new swarm-native companies?
3) What becomes scarce as swarm capabilities themselves commoditize (through frontier models becoming better at swarming, swarm capabilities built into harnesses, etc)? Cultures and teams that can define reward functions well and otherwise get out of the way?
Read the Stand with Jensen letter and consider adding your name. #standwithjensen #bringjensenback standwithjensen.com/
Man do you guys ever feel like you’re living in a different world from everyone around you? It’s this vertigo of seeing so so much and being utterly unable to explain it .. my wife, my friends, even many of those who are interested haven't yet discovered how freakishly deep this rabbit hole goes and all the fascinating implications.
One thing that’s changed is I am OBSESSED with my computer in a way that I haven’t been since we got our first 286. I can’t get enough of it and being away from it creates the kind of fomo that I used to feel for things in the “real world”. When I’m away, I catch myself constantly thinking about the experiments and agents I’ve left running in the background.
And then when stuff goes right, really right, there is this visceral dopamine hit that I live for. It’s variable schedule rewards ish but way more satisfying than any other casino game because many wins are the baseline for other wins. And most of the wins are more about satisfying curiosity or learning through experimenting than metrics or utility, but who cares.Â
The other thing that is so interesting is that we get to go back to all the discussions of how we’ve organized society and the economy and revisit every assumption. Hayek, Burnham, Adam Smith, W Brian Arthur, Jevons … we can model all of their ideas and worlds with agents, we can think about the interplay between humans and machines in these frameworks, and we can ask ourselves which ideas are the winners and losers in a world of economic convexities (i.e. the marginal cost of labor is zero, intelligence goes from scarce to abundant).
Anyways just a crazy crazy time that is intellectually nourishing in a way I’ve never experienced. That's my dispatch for the day from the psychosis frontier .. what a cool moment to be a part of.
Anish Acharya retweeted
wrote up some thoughts on ai x gaming with @illscience and @Tocelot.
read more below!
Coding agents are going to radically change the economics of game development, but the much more interesting idea is that consumption itself becomes a creation primitive. What happens when playing a game and making a game become the same activity? When anyone can modify the rules, invent new mechanics, or extend a world simply by asking an agent to do it? I suspect millions of people will build games that nobody else ever plays, just for the joy of creating them.
@cdixon makes a great point in Read Write Own about how differently music and gaming responded to the internet. Music largely fought fan participation while gaming embraced it, and some of the biggest franchises in the world emerged from mods.
Models take this idea to an extreme. Games become living, malleable worlds rather than finished products, and the distinction between creators and consumers starts to disappear.
We spent this episode talking about everything from resurrecting forgotten games to compressing AAA development cycles to entirely new genres that wouldn’t have been economically viable before. But perhaps the biggest opportunity is building worlds where the creativity of millions of players can compound into something far more interesting than any individual developer could have imagined.
AI is making games faster and cheaper to build, giving developers more room to take creative risks, revive overlooked genres, and let players reshape the worlds they play in.
In this conversation, a16z's Jon Lai and Anish Acharya join Elena Burger to discuss why human taste still determines what makes a game fun, how modding created some of gaming's biggest franchises, the opportunity for new distribution platforms, and games that adapt to whatever players choose to do.
0:00 Intro
0:57 How AI can now build a game
6:21 What shorter dev cycles unlock
10:08 Can AI make a game fun?
12:27 When creating is the game
14:01 Building a Habbo Hotel record shop
16:16 His son rebuilt Ghost of Tsushima in Minecraft
17:54 Mods: where billion-dollar franchises come from
20:36 Decompiling and reviving dead games
22:45 The renaissance of niche genres
25:51 Managing AI agents like an RTS game
28:02 The business of AI games
33:53 What excites and scares studios
35:16 The never-ending game
37:56 The games they'd finally buQild
YouTube: youtu.be/HDRy6XmE2RQ
@VirtualElena @illscience @Tocelot
the mythos 2 max price per discovered vulnerability is 4x higher for an incremental 1% discovery rate, knowing that the buyer is utterly price inelastic - the value of that 1% could be the value of your entire company
note the x axis is log scale
Absolutely love this story. There will be emergent methods for everything from chess to product management to commerce that will be totally non-intuitive, because they were devised by an alien mind.
Some will be discovered through parallelization (brute force) but others will surely come from entropy or extreme specialization + comparative advantage.
Breaking the exploration / exploitation tradeoff in many domains has tremendous implications for the world.
JOE ROGAN: The computing power on your phone today. can you beat your phone at the highest level?
MAGNUS CARLSEN: No. No chance.
JOE ROGAN: That’s crazy. Wasn’t Deep Blue basically a whole stack of computers? And your phone is probably more powerful than that today.
MAGNUS CARLSEN: It’s shocking. I have no chance against my phone.
JOE ROGAN: So how do you even play against computers?
MAGNUS CARLSEN: I rarely play against engines because they just make me feel stupid and useless. I think of them more as a tool.
JOE ROGAN: But have computers actually changed how you understand chess?
MAGNUS CARLSEN: Absolutely. Neural networks have improved our understanding of the game immensely.
JOE ROGAN: What did you learn from them that humans weren’t already doing?
MAGNUS CARLSEN: One big thing was long-term king safety. Humans used to push pawns mainly as an attacking tool. But the engines showed that you could push a pawn and create a weakness 20 or 30 moves later that would make the opponent’s king less safe.
JOE ROGAN: So the computer basically showed you ideas that humans weren’t seeing.
MAGNUS CARLSEN: Exactly. There was a period where you could clearly see which players were using these new neural-network ideas and which weren’t.
JOE ROGAN: And you were one of the guys who figured it out early?
MAGNUS CARLSEN: Yeah. My coach got into it very quickly, and we gained an advantage over almost everybody.
JOE ROGAN: That’s insane. The machines aren’t just better at calculating, they’re actually teaching the best players how to think differently.
MAGNUS CARLSEN: That’s probably the biggest impact they’ve had on chess.
I try to use every coding model but have a hard time with those that don't have a dedicated s-tier harness. I would call codex / claude code and cursor s-tier. I use opencode as well, though not as heavily.
This has held me back a little bit from really taking Argon and Wonder for a test drive - it's hard to see what the models are capable of when you're missing a lot of the quality of life basics (robust remote control, browser control). And as the best is increasingly defined by capabilities like agent teams, the great harness+model pairs will diverge further and further from the sum-of-parts pairs.
I've heard that the best models are post-trained alongside their harnesses, and that those harnesses have more information than a vanilla harness (preserved thinking / CoT etc) - partially to avoid distillation but more importantly to capture the economic benefits of vertical integration.
Finally it seems like there's a whole emerging area of "getting the most out of the models" at the harness/agent layer. Admittedly some of this is vibes based, but the right harness + model pair is like living in the future.
General Partner at Andreessen Horowitz Anish Acharya (@illscience) says using AI to edit the weights of other models is a bigger deal than people realise and could unlock continuous learning:
"There's actually two huge breakthroughs that have happened in the last few months that I think are under discussed... [One is] the paper that was released on Monday about having models edit the weights of other models, which gives you the sort of continuous learning opportunity."
"It was released and there was enthusiasm and then it was criticised, but I think it's a bigger deal than people appreciate."
"The smartest people I've been talking to have been able to not just replicate, but extend it."
"Now look, there may be some constraints, like there may be opposing ideas and models can't hold both of them in their head. You can't be highly conscientious and highly open, right? Because that sort of implies that the same weights are pointed in different directions."
"But you may be able to do things like get models that are able to have high entropy in non-verifiable domains, like beautiful art and prose, but then high precision in verifiable domains where they can also write great code and be mathematical."
"And we've never been able to reconcile those trade-offs before."
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! đź§µ
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Anish Acharya retweeted
.@illscience General Partner at @a16z was one of our very first guests on the show
(which I am so grateful for).
But his quote on product and marketing, keeps coming back to me which every founder should learn from.
Thank you @illscience for all your support!!
Our podcast show is one year old!
2M+ downloads so far and counting!
Most quit after the first few episodes but we are still here and going strong!!
A huge thank you to all our previous guests, partners and friends of the channel for all your support so far!
We have one mission: Help you learn from the most brilliant minds shaping AI!
Back to the microphone in 10 minutes!
Congratulations to Vincent and the team on a really thoughtful product launch.
Quest feels like it’s pointed in a different direction than other personal agents, with more of an emphasis on exploring and extending your personal interests. More "personal" than "agent" perhaps.
What’s clear is that personal agents represent a new way of using computers and intelligence that consumers are really into. It feels like there’s a wide product surface to explore, and I’m excited to see more efforts pointed in different directions like this.
Introducing Quest, the personal agent that studies you like a profiler.
Quest is for believers-only. It requires radical transparency: full access to your life so your agent can craft your memory.
If you give it a shot, you’ll unlock a personal advisor, always in your corner, helping you and encouraging you to start your quests.
Quest is built different:
1. Memory from day one: 100+ wiki files created on you, all your relationships mapped.
2. Phone-Use: Quest has its own Android phone, and can control messaging apps like WhatsApp, Instagram or TikTok.
3. Fluid UI: Quest builds tailor-made iOS mini-apps around each of your quests.
4. Mothership of AI Assistants: Quest can manage your Instinct and Muse on your behalf.
5. Self-Improving Proactivity Harness: Quest reacts to anything, not just your email, and its proactivity evolves from your feedback and its own mistakes.
Btw here is a link to the record which is dope as hell: juno.co.uk/products/cpen-hyp…
5
619
my latest thinking on personal agents (from the pod w @davidpawlan + @joshelman):
winners / losers
- there's so much discussion of structural losers from personal agents (delta / amzn / comcast) and very little about the relative winners
- presumably the long tail of merchants with great products and mediocre marketing benefit a lot
- the idea that agents will perfectly mirror their principals is a fantasy imo... agents will have revealed non-rational preferences (i.e. for clean API surfaces) that are those of an alien mind
- implied is a consumer research / wirecutter product where agents research and share their experiences with each other
agent adoption
- we should be giving agents laws and principles vs rules
- fears of embarrassment will hold more people back than fears of privacy
- agents need to be designed so some implicit preferences stay implicit, so their human can save face
- group chats have an implicit social contract, and agents are only part of that contract if they're present from day one
- agents can create a level of social indirection that actually brings people closer together (i.e. what tomo is doing with groups of humans who share aspirations)
agent networks
- if muse / instinct / wabi is consumer openclaw, what is consumer moltbook?
- what economic exchanges can agents make on behalf of their humans?
- is there a concept of "free" agents that operate within these networks of their own volition?
David Pawlan, creator of Assistant Benchmark, has tracked 147 AI assistants so far, and in his group chats of AI power users, the number one fear is trust. It comes up in 22% of roughly 3,600 messages.
He puts agents through head-to-head tests, like ordering same-day flowers as cheaply as possible or planning a weekend in LA that's both luxurious and cost-efficient, to see which ones can hold competing preferences.
He joins a16z's Anish Acharya and Josh Elman to discuss why the scariest agent failures are the embarrassing ones, why agents are where e-commerce was in 1997, why scale is still a moat, and the "dark agent" that could speak to the parts of ourselves we don't reveal.
0:52 Inside a meetup for 100 agent nerds
2:52 Why we'll all have an army of agents
5:36 Trust is power users' top fear
6:28 Building fences around your agent
10:14 Privacy loses to free pizza
12:41 Can agents balance price and speed?
14:19 How agents define luxury differently
19:44 Migrating from one agent to another
21:06 What your photo roll tells an agent
25:48 Will agents make us more rational?
27:38 Why paternalism kills products
31:43 Gossipy agents as social glue
33:41 Scale is still a software moat
36:55 From OpenClaw to agents hanging out
39:23 Agents that introduce you to people
46:15 Predictions and the dark agent idea
YouTube: youtu.be/dSiTwPl0t1U
@DavidPawlan @illscience @joshelman
Anish Acharya retweeted
The new Oboe app was a lightbulb moment for me in how AI games, movies, etc. get huge
It’s infinite puzzles where the user designs their experience - but curated, so the opposite of a slop feed
It’s one of the only AI mobile apps stealing screen time from incumbents for me 👀
We’ve pivoted Oboe. The first version, which we spent the past two years building, didn’t work. No excuses. Here’s the full story, culminating in what comes next: an all-new Oboe, available today.
We launched Oboe with a mission to make humanity smarter. That mission is core to who we are, and it hasn’t changed. The first way we thought to achieve it was by leveraging frontier models to offer everyone personalized tutoring.
But through many iterations over the past two years, it became clear that we built the wrong product. First, we’d made strategic decisions that left us undifferentiated in the age of LLMs. Second, by focusing on utility above all else, there was little organic incentive for people to return to the platform. And most importantly, no matter how great our product was, not many people were motivated to actively learn every day.
Despite our best efforts to serve everyone on the planet, we landed on a small, loyal user base that loved a product which wasn’t growing fast enough.
So we took a step back: How else could we serve everyone on the planet and still achieve our mission?
We decided instead to focus on a motivation everyone has: to have fun. Every day, billions of people entertain themselves on their mobile devices, filling countless minutes with mindless scrolling and brainrot that doesn’t make us smarter. In many cases, it actually makes us stupider.
What if we could replace scrolling that makes us stupider with scrolling that is both fun and makes us smarter? That’s the new Oboe.
It’s a fully personalized feed of dynamically generated, interactive, anti-brainrot challenges. Every scroll is designed specifically for you, to help you get smarter in a few minutes every single day.
We’ve been blown away by usage from our thousands of beta users. The differences in engagement and retention between this and the original version are stark. The product is working. It’s already clear that this new direction taps into a universal need that none of our previous iterations did.
For all of those people out there like me who feel like doomscrolling wastes their precious time, who enjoy feeling fulfilled, and who wish there was a better alternative: this one’s for you.
We can’t wait to see what you think.
David Pawlan, creator of Assistant Benchmark, has tracked 147 AI assistants so far, and in his group chats of AI power users, the number one fear is trust. It comes up in 22% of roughly 3,600 messages.
He puts agents through head-to-head tests, like ordering same-day flowers as cheaply as possible or planning a weekend in LA that's both luxurious and cost-efficient, to see which ones can hold competing preferences.
He joins a16z's Anish Acharya and Josh Elman to discuss why the scariest agent failures are the embarrassing ones, why agents are where e-commerce was in 1997, why scale is still a moat, and the "dark agent" that could speak to the parts of ourselves we don't reveal.
0:52 Inside a meetup for 100 agent nerds
2:52 Why we'll all have an army of agents
5:36 Trust is power users' top fear
6:28 Building fences around your agent
10:14 Privacy loses to free pizza
12:41 Can agents balance price and speed?
14:19 How agents define luxury differently
19:44 Migrating from one agent to another
21:06 What your photo roll tells an agent
25:48 Will agents make us more rational?
27:38 Why paternalism kills products
31:43 Gossipy agents as social glue
33:41 Scale is still a software moat
36:55 From OpenClaw to agents hanging out
39:23 Agents that introduce you to people
46:15 Predictions and the dark agent idea
YouTube: youtu.be/dSiTwPl0t1U
@DavidPawlan @illscience @joshelman




























