
Founder @isotopes_ai Past life: CTO, @scale_ai, CPO, @cloudera. Co-Founder/CPO, @Hortonworks. Engineer sheepdog. Self-confessed old soul.
Never still, mostly move ahead
Joined May 2009
- Tweets12.9K
- Following238
- Followers17.6K
- Likes2K
Arun C Murthy retweeted
Open weights are having a moment. Better models, cleaner licenses, more ways to tune and serve.
Exploring Inkling @thinkymachines for a focused loop in Neocortex raised a simple question. Do you really get the same experience?
isotopes.ai/blog/everyone-is…
Arun C Murthy retweeted
Most AI analytics tools answer questions.
The harder question is: what do they actually know, how do they know it, and what happens when they are wrong?
That’s why we built Neocortex, the self-learning model behind aidnn.
Read more:
isotopes.ai/blog/neocortex-s…
In a world with abundant intelligence, every enterprise runs on two kinds of capital: the human capital of its people's judgment, and the cognitive capital of the reasoning ability it builds and owns about the business. The second is new, and it is where durable advantage accrues.
White-collar work in the enterprise forms a natural hierarchy.
→ At the base sit the *systems of record*: the warehouses, lakes, ERPs, and CRMs that hold the data.
→ Above it comes *knowledge work*: the assembly, reconciliation, and verification that turn scattered records into a trustworthy answer. It is procedural even when it's hard, and it's staffed today by armies of analysts and outsourced teams.
→ At the top comes *judgment work*: weighing trade-offs when systems disagree, owning the answer when challenged, the part that was always the key for world class decisions.
Software increasingly absorbs the knowledge work. It still needs the judgment work. Cognitive capital accrues from both knowledge work and cognitive work.
Today @isotopes_ai is introducing *Neocortex*, the self-learning system behind aidnn. Each answer it produces draws on a set of experts from your organization's verified knowledge, and shows you every expert who contributed and every belief it relied on. We built it the way the brain learns, by doing rather than by being instructed:
→ It holds what your organization believes
→ It learns how your organization does the work
→ It judges that work and learns from how you judge it
→ It consolidates every answer into durable cognitive capital
A critical difference from existing approaches: Neocortex learns from doing the work, not from instructions you write down for every task. Every correction shows it how the work should be done, and the one thing it does better than any tool you've used is this: it never needs to be told twice. The verification layer within aidnn lets Neocortex *learn* the nuances that matter and, just as important, *unlearn* them as business realities change.
The payoff: the reasoning your best people do today stops evaporating after each question and starts compounding into an asset you own.
Interested in learning more?
We are happy to show you Neocortex with *your* data:
isotopes.ai/contact-us?sourc…
Arun C Murthy retweeted
The work can now speak for itself.
The ability to send your boss a session and have the agent answer "Why did you assume that?" instead of pinging you back on slack required making agents play multi-player.
Most AI products give you a chatbot. You ask, it answers. But that doesn’t work for serious analytical work.
Your team doesn't watch every step. They start an analysis, go to meetings, come back the next day.
Colleagues need to understand what happened.
Sometimes you need to rewind and try a different path.
A chatbot can't do any of that.
So we threw out the chatbot model and rebuilt around one idea: the event stream.
Every action aidnn takes is emitted as a structured, ordered event. That single decision unlocks everything:
→ Teammates jump in, ask questions anchored to specific moments, and keep going
→ Close your laptop, come back tomorrow — zero lost progress
→ Branch from any checkpoint and compare results
→ See assumptions and reasoning before execution, not after
The result feels less like querying a black box and more like working with a teammate who documents their work and can be reviewed for methods, not just results.
Full breakdown on the @isotopes_ai blog:
blog.isotopes.ai/building-a-…
Arun C Murthy retweeted
Specialized agent teams organized to achieve clear goals, utilizing a remote code executor to separate data transformations from reasoning. Agents execute code remotely, returning only relevant summaries, thereby maintaining a separation between perception and execution.
We recently talked about the "Team of Rivals" architecture which we bring to bear @isotopes_ai to scale sophisticated AI Agents, here is our paper on @arxiv:
arxiv.org/abs/2601.14351
Discuss: news.ycombinator.com/item?id…
Arun C Murthy retweeted
We just published our whitepaper on building production-grade agentic AI.
Core thesis: analytics isn’t a chatbot problem - it’s an orchestration problem.
Get the whitepaper → isotopes.ai/whitepapers/orch…
We just published our latest whitepaper "The Orchestration Imperative for AI Agents" and this one is especially interesting for leaders who want to learn about the next evolution of AI.
Here's the core insight: Production-grade AI Agents aren't a chatbot problem. They are an orchestration problem.
Most multi-agent AI systems fail for predictable reasons—coordination overhead scales nonlinearly, context gets contaminated and rots progressively, and debugging becomes impossible.
We've watched teams spend months building "agentic" solutions only to hit these walls.
This whitepaper distills the principles @isotopes_ai uses to make sophisticated AI agents work in production: structured workflows, plans as contracts, hierarchy that tames complexity, and humans owning the "what" while the system handles the "how", and more.
If you're evaluating any agentic AI for your business - or building one - this is definitely worth a read!
Get the whitepaper to learn more → isotopes.ai/whitepapers/orch…
Arun C Murthy retweeted
This idea that "AI models will not keep scaling" was conjecture in Nov 2024 when we wrote the first prototype.
Through 2025, we have built a system which does not depend on context windows growing and lossy compactions, but go faster with lots of fast flash LLMs
Orchestrating a Team of Rivals
We @isotopes_ai just published our in-depth technical paper on how aidnn achieves production-grade reliability!
Model scaling is slowing down. But AI impact doesn't have to plateau.
The path forward isn't faster clock-speeds. It's multi-core. Instead of waiting for a single smarter model, aidnn orchestrates multiple LLMs with opposing incentives.
No business relies on one employee to handle critical operations. We shouldn't architect AI systems around single-agent execution either.
aidnn deploys 50+ specialized agents organized as specialized teams:
* Planners create execution strategies
* Executors perform the work
* Critics validate outputs with veto authority
* Remote code execution keeps data isolated from reasoning models
Each of these agents are coding agents resulting in a highly repeatable & reliable agentic system.
We call it "orchestrating a team of rivals." Multiple agents with different failure modes catch what individual models miss. Hierarchical veto authority prevents errors from propagating.
The result: 90%+ error prevention through layered validation. And your data stays secure.
This is the architecture behind financial close automation at Invisible Technologies and others — turning weeks of manual work into hours of validated analysis.
Read the full technical paper here:
isotopes.ai/technology
Arun C Murthy retweeted
architecture behind aidnn: agentic teams with opposing incentives validating each other’s work.
This is how we’re getting production-grade reliability for financial analysis.
Read more → isotopes.ai/technology
Orchestrating a Team of Rivals
We @isotopes_ai just published our in-depth technical paper on how aidnn achieves production-grade reliability!
Model scaling is slowing down. But AI impact doesn't have to plateau.
The path forward isn't faster clock-speeds. It's multi-core. Instead of waiting for a single smarter model, aidnn orchestrates multiple LLMs with opposing incentives.
No business relies on one employee to handle critical operations. We shouldn't architect AI systems around single-agent execution either.
aidnn deploys 50+ specialized agents organized as specialized teams:
* Planners create execution strategies
* Executors perform the work
* Critics validate outputs with veto authority
* Remote code execution keeps data isolated from reasoning models
Each of these agents are coding agents resulting in a highly repeatable & reliable agentic system.
We call it "orchestrating a team of rivals." Multiple agents with different failure modes catch what individual models miss. Hierarchical veto authority prevents errors from propagating.
The result: 90%+ error prevention through layered validation. And your data stays secure.
This is the architecture behind financial close automation at Invisible Technologies and others — turning weeks of manual work into hours of validated analysis.
Read the full technical paper here:
isotopes.ai/technology
Arun C Murthy retweeted
Exactly! One analyst spent 3 hours on Q3 working capital analysis. With aidnn they can just share a notebook, CFO clicks explain on anything they want, VP Finance branches it to test scenarios. Done in 20 minutes...
Arun C Murthy retweeted
People really get it when they see branching. You share monthly expense analysis. Someone highlights a weird AWS spike, clicks "Explain," gets the breakdown. Then thinks "what if we switched regions?" and branches to test it. That's the moment!
Recently, @far33d from @a16z nailed it: AI 1.0 gave individuals superpowers. AI 2.0 will give teams superpowers.
With aidnn from @isotopes_ai you can live in the future.
We built aidnn to be multiplayer with shared Organizational Memory from day one.
Everyone sees the same context, numbers, assumptions, and process - updated in real time.
While others are trying to make single-player AI tools work for individuals, we're already delivering multiplayer intelligence with collaboration, shared history, and organizational memory built in - today.
Arun C Murthy retweeted
Your analytics stack just became AI-assisted. Starting today, we @isotopes_ai opened aidnn to everyone!
Teams at @InvTechInc and others are automating their analytics—KPI refreshes, margin analysis, anomaly detection, board prep.
Not a BI dashboard. Not ChatGPT. A true agentic platform specifically for analytics.
isotopes.ai/signup
Arun C Murthy retweeted
Flight attendant comes over intercom: “your flight is oversold, I need two volunteers to take a flight this afternoon for a $300 voucher.”
I spring into action. I was born for these moments. This is why I studied game theory. Literal tingles. I pull out my megaphone and address our boarding gate: “guys, they absolutely must off board at least two passengers or this bird isn’t taking off. Every 3 minutes they will increase their bid by $200. If we all hold out for 45 more minutes, 2 of us can make nearly 5 grand each. Textbook prisoners dilemma, this is (3,3) in its most visceral representation. I’m not even going to give up my ticket but I know value when I see value. Let’s make Nash proud ladies and gentlemen. Do not acquiesce. Do not defect.”
Anyhow 2 people immediately accepted $500 and I called them retards on my megaphone and was placed in a detention room by airport police. This is why we study game theory.
Arun C Murthy retweeted
The Espresso Revolution
Over the past few months, almost everything we know about espresso brewing has been called into question. Today, I will share with you the story of espresso, the meaning of SOUP, and the radical implications of GRINDING COARSER.
------------------------------
Espresso is traditionally prepared with finely ground coffee brewed under pressure using an espresso machine. It is almost impossible to make well. You need, at minimum, hundreds of dollars of expensive coffee equipment, dozens of hours of practice, and great coffee beans. But it is so finicky that each day - and often throughout the day - one must adjust several parameters to maintain a consistent flavor profile. To be honest, I have had very few good espressos in my life – they are almost always defective. On his blog, @jimseven famously posited the question to cafe operators, "what percent of the time you order an espresso at your own cafe do you expect it to taste good?" The answers are abysmally low. So how did we get here?
The Single Shot
Traditional Italian espresso followed a simple formula: seven grams of coffee, finely ground, extracted at nine bars of pressure for twenty-five seconds, yielding about twenty milliliters of concentrated beverage. The ~1:2 ratio became gospel, codified by Andrea Illy in the 1990s as the definitive standard. This was espresso as most of the world knew it—dark roasts, consistent technique, that distinctive golden stream pouring from the portafilter like honey, crowned with that layer of reddish-brown foam we call crema.
The equipment was straightforward: a pump-driven machine applying constant pressure, a good burr grinder set fine, and muscle memory developed through thousands of pulls. This was espresso. This was the foundation.
The God Shot
In the 2000s, web forums like Home-Barista.com and Coffeed brought nerds from around the world together to experiment with coffee. And experiment they did. Using digital scales (imagine weighing your coffee!), freshly ground coffee (not from a dosing container!), and "bottomless" portafilters, home enthusiasts staked a claim in a new way of making espresso. Enthusiasts would pilgrimage David Schomer's legendary Espresso Vivace to taste the platonic ideal of coffee: "the God Shot."
In one sense, the God Shot era was the wild west of coffee. Pioneered by "home baristas" without formal training, espresso recipes developed some weird quirks. John Weiss started using a hypodermic needle to de-clump grounds in his portafilter, in what became known as the Weiss Distribution Technique (WDT). David Ross started spraying coffee beans before grinding – we now call this the RDT. Recipes centered around 2:1 double-shot ratios in naked portafilters, with 18.5 grams in, 37 grams out in 28 seconds. The lucky few got their hands on the latest tech from VST: expensive new portafilter baskets and groundbreaking coffee refractometers. But the era was defined by a general sense of yearning for what one knew could be better.
As Tim Wendelboe's dramatic style of light roasting became more popular, these "god shots" tasted worse and worse, and the old rules weren't working anymore. The specialty industry was too focused on the latest La Marzocco model to develop new recipes, and espresso stagnated in flavor. If you ordered espresso in a cafe in 2015, and you'd be lucky to taste something palatable.
Something had to change.
Allongé
Scott Rao saw the problem clearly. In the late 2010s, working with John Buckman and the Decent Espresso machine, Rao debuted what he called "allongé"—a style of very long espresso designed specifically for lightly roasted coffee. The approach was counterintuitive: grind even finer than traditional espresso, use pressure profiling to manage flow, and push water through the puck at ratios of 1:3 or even 1:4. Pull fifty to sixty grams of liquid from twenty grams of coffee. The extraction times stretched past a minute.
The results were remarkable. Rao achieved extraction yields that bordered on the impossible–24%, even up to 29% on some Kenyan coffees. The Decent Espresso machine made it possible through precise flow control and pressure profiling, but the technique demanded immaculate puck preparation. Any channeling, any uneven distribution, and the shot fell apart. Needle distribution tools became essential. Deep WDT wasn't optional.
Allongé worked beautifully for certain coffees—those bright, high-altitude African lots in particular. The message was clear: more extraction was better. Push the boundaries. Pull more from the coffee. If you could hit 25%, why not try for 27%? The entire trajectory of modern espresso pointed in one direction: grinding finer, increasing yield, and pushing flavor further than we ever tasted. And then came the scientists.
Turbo Shot
In January 2020, a paper landed in the journal Matter that seemed to chart a new course. "Dr. Coffee" @chhendon published "Systematically Improving Espresso: Insights from Mathematical Modeling and Experiment," bringing mathematical rigor to what had been intuitive adjustments.
The turbo shot approach was elegant: grind coarser, not finer. Use less coffee. Lower the pressure to six bars. Pull the shot fast–seven to fifteen seconds fast. The mathematical models showed that fine grinding created problems. Fine particles clogged the puck, causing channeling and uneven extraction. Some parts of the puck over-extracted while others under-extracted, producing muddy, inconsistent results.
By grinding coarser and extracting quickly at lower pressure, turbo shots achieved higher average extraction yields with better consistency. They ran fast but tasted clean. The shots had clarity but less body, brightness without astringency. The paper was a triumph of systematic thinking, giving the community scientific backing for what some had discovered through experimentation.
The science kept going. Famed coffee blogger Jonathan Gagné of Coffee Ad Astra continued characterizing espresso dynamics quantitatively, in what's culminated in this year's newly published The Physics of Espresso. Gagné synthesized the new understanding: modern baskets with different geometries fundamentally changed flow patterns. Needle distribution tools weren't fussy accessories—they were essential for even extraction. Paper filters could prevent fines migration. Grinders with cleaner particle distributions required different approaches than traditional espresso grinders. The community finally had the framework to understand why techniques worked.
The trajectory seemed clear. Better understanding led to better technique. Higher extractions, more consistency, more sophisticated equipment. The future of espresso was professional, optimized, increasingly technical.
And then SOUP happened.
SOUP (Spro Only Un-Pressurized)
This year, something emerged from the Espresso Aficionados Discord that shouldn't have worked. Coffee industry legend Bailey Manson noticed a weird new brewer from OXO – yes that OXO, the vegetable peeler company – called the OXO Rapid Brewer. This was a forty-dollar manual coffee maker designed for travel that looked like a clunky and awkward take on an Aeropress. The OXO wasn't even marketed as an espresso machine. It was barely taken seriously.
But Bailey discovered something that broke every rule we'd just learned. By grinding coarse–somewhere between pour over and espresso–and using paper filters above and below the puck, the OXO could produce something intriguing. Some form of coffee came out of this strange device, between ~7-8% TDS with pretty respectable extraction yields ranging from 19-20%. The strangest phenomenon was that the OXO had very little pressure; its integrated pump release valve kept the brew chamber well below one bar.
Lance Hedrick picked one up, tried a few brews, and loved it. He coined the term "SOUP" to describe these unique shots: Spro Only Un-Pressurized. And in just weeks, the OXO spread like wildfire.
The first time I tried SOUP I was genuinely blown away. In just 15 seconds, with a $40 plastic brewer, I produced one of the most memorable coffees of my life. Hydrangea's Letty Bermudez Geisha tasted more floral, with better balance and body. Because of the low pressure, the coffee had zero astringency. The lower lead to unbelievable clarity. All of this added up into some of the most complex coffee I've ever tasted.
This shouldn't have been possible. This random coffee was better than the high-extraction allongés, better than perfectly dialed turbo shots, better than traditional espresso pulled on five-thousand-dollar machines. And without any fancy equipment, I got it right on the very first time.
The Great Unknown
SOUP isn't a continuation of recent espresso trends. It's the opposite of everything we'd been learning.
Allongé told us to grind finer and extract higher—SOUP grinds coarse and extracts lower. Turbo shots used mathematical models to optimize consistency—SOUP achieves consistency through a method so forgiving it borders on crude. The entire trajectory of modern espresso pointed toward higher extraction yields, and here was SOUP, stopping at 19-20% and tasting better than shots extracted to 25%.
It would have been crazy to tell someone in 2020 that the best-tasting espresso would come from extraction yields under 20%. That contradicted everything. Rao's allongés demonstrated that light roasts needed high extraction. Hendon's paper showed that proper technique could push yields higher while maintaining clarity. Gagné's work gave us the tools to understand and optimize every variable.
And yet SOUP–made with a travel brewer, coarse grinds, barely any pressure, and moderate extraction–tastes better.
What do we actually know about brewing coffee?
The most unsettling thing about SOUP isn't that it works. It's that we don't fully understand why it works. The conventional wisdom said you needed pressure to extract espresso-level intensity. SOUP uses essentially none. The science said higher extraction yields produced better flavor from light roasts. SOUP stops well short and tastes cleaner, sweeter, brighter. The equipment manufacturers told us we needed precise temperature control, flow profiling, pressure ramping. SOUP uses a manual plunger and boiling water.
Everything we learned over the past five years–the sophisticated equipment, the mathematical models, the optimization of every variable–feels suddenly questionable. Not wrong, exactly. Just... incomplete. Like we'd been building an increasingly complex edifice on a foundation we didn't fully understand.
The SOUP technique itself is almost absurdly simple:
1. Grind 22 grams of coffee about Aeropress fine.
2. Add paper filters above and below the puck to prevent fines migration.
3. Add 82 grams of boiling water.
4. Put on the lid and pump.
And that's it. In about 10-15 seconds, you get a pale brown liquid with no crema–a haunted ghost of espresso that looks like one of the worst Americanos you've seen–at about a 7-8% TDS.
It's so forgiving that people are getting great results on their first try. No dialing in. No channeling issues. No need for a $3,000 grinder or a $5,000 espresso machine. The barrier to entry is forty dollars and a willingness to question everything you thought you knew.
What Now?
SOUP poses some fundamental questions to the specialty coffee world. How deep of a local optimum are we in? How much better can coffee taste? And what are the actual primitives of extraction that contribute to flavor?
The second order effects are dramatic. Coffee brewed at 19% EYs will have to cost more than those at 23%. As of this writing, there are ~no commercial espresso machines that can reliably produce SOUP, so why would you go out to buy a coffee worse than the one you can make at home faster? If you only need to spend $40 on coffee equipment, what happens to the accessories market?
I know of no comparable enthusiast community where something radically cheaper turned out to be radically better. HiFi audio, camera gear, music equipment, kitchen knives... all these rabbit holes get more expensive. What if the last stop becomes the cheapest?
In the last 20 years, there was a sense of linear progress. Golden Honey. The God Shot. Allongés. Turbo Shots. But now a technique that violates almost every principle we've established is producing some of the best coffee anyone's tasted.
That sometimes the only way forward is to question everything we think we know and start over.
Grind coarser. Build less pressure. Accept that the numbers might lie and your taste buds might be right. The future of espresso is uncertain.
And that's exactly what makes it exciting.
Arun C Murthy retweeted
we would never give out headphones
The future of travelling with @ryanair and if by any chance your foot size is more than 10/11 an extra 15% of the ticket price will be charged to you.
Arun C Murthy retweeted
In med school, I learned that Huntington's disease was one of the most devastating inherited conditions. It caused progressive neurological decline, starting around age 30. There was no treatment. Now, Huntington's has been treated for first time.
This is why we fund science.
Arun C Murthy retweeted
Overwhelmed by the reception, thank you for all the kindness and love!
Here is a quick tour of aidnn and the awesomeness there-in:
youtu.be/8pz7Zi0Id_M
Interested in more action? See the demos: isotopes.ai/#demos
Arun C Murthy retweeted
🚀 Big News in AI & Big Data!
Isotopes AI has officially emerged from stealth with a massive $20M seed round, introducing its AI agent Aidnn — designed to solve one of big data’s oldest challenges: making insights accessible to business leaders without deep technical skills.
💡 Instead of relying on data engineers, managers can now ask questions in natural language and Aidnn will fetch, clean, and analyze data from multiple sources like ERP, CRM, finance apps, and cloud storage.
What makes Isotopes stand out?
✨ Founded by ex-Hortonworks & Scale AI leaders
✨ Applied for 10 patents already
✨ Aidnn can draft complex reports, show reasoning steps, spot anomalies & make recommendations
✨ Works securely without sharing enterprise data with AI model providers
🔹 Seed round led by NTTVC’s Vab Goel
🔹 Competing with giants like Salesforce’s Tableau and startups like WisdomAI
📊 The company says Aidnn is “far beyond a simple chatbot” — it’s built to transform enterprise decision-making with trustworthy, contextual insights.
#TheTechFin #USA #Europe #Canada #Tesla #xAI #ElonMusk #AI #Robotics #Innovation #TechNews #EV #Shareholders #FutureTech 🚗🤖
Arun C Murthy retweeted
Then, head over to isotopes.ai/signup.html to get a free account to play with aidnn for our initial phase!
Fair warning: We are under water, and are slowly getting through all the signups - appreciate your patience!
Arun C Murthy retweeted
Ooo...emmm...jeee!
So, thankful - We are overwhelmed with the feedback for aidnn and the love showered on the team and the product!
isotopes.ai/#demos has a bunch of demos you can click thorugh to get a sense of how aidnn fits in as a "co-worker" and the complex tasks aidnn can handle.





















