My retweetledi
My
2.4K posts
My retweetledi
TPN OTC Deal Transparency Disclosure.
We’re excited to share that @const_reborn has acquired 67,500 TPN alpha tokens via OTC.
The transaction took place on Mar 25, 2026 at 3:45:24 PM (UTC) for 301 TAO and is fully verifiable on-chain.
Thank you Const, this vote of confidence means a great deal to the entire team and motivates us to keep building toward the future of the subnet.
English
My retweetledi
For context: Karpathy is a former head of AI at Tesla, co-founder of OpenAI, Stanford PhD.
One of the most recognized AI researchers on the planet.
He laid out a vision for "Auto Research" on a major podcast @NoPriorsPod
That vision?
Crowdsourced algorithm optimistion with proof of work and a collaborative network.
Sound familiar?
x.com/tigfoundation/…
English
My retweetledi
One of the most famous AI researchers on the planet just described his "big idea" on a major podcast.
He had no idea someone already built it. Here's everything from last Friday's TIG community call
English
My retweetledi
@ReadOnlyUp @GoogleResearch @tigfoundation google drops a 6x compression algo and validates the exact market TIG is building for
decentralized SOTA algorithms just got their biggest proof of concept from mountain view
English
My retweetledi
My retweetledi
Hi Alex
I believe the biggest leverage in know-how would be in mathematics because of the application to AI-assisted algorithm discovery (AlphaEvolve, etc).
See here for details x.com/Dr_JohnFletche…
Regarding trying different data types, yes, I feel it would be healthy to have a diversity of approaches.
That's why its odd that there's not a single company, startup or hyper scaler, talking about capturing and utilising know-how.
John Fletcher (𝔦, 𝔦)@Dr_JohnFletcher
English
My retweetledi
Interesting how the most recent advancement in AI was due to a new advanced algorithm rather than increased compute. Which is why @tigfoundation $TIG is still a very relevant project, as it is the only one working on decentralized incentives towards improving algos.
English
My retweetledi
Multi-billion dollar investment rounds continue to be raised based on the following assumption..
"In key domains such as mathematics, coding, and science, the knowledge extracted from human data is rapidly approaching a limit. The majority of high-quality data sources — those that can actually improve a strong agent's performance — have either already been, or soon will be consumed."
Quote from "Welcome to the Era Of Experience" google.com/search?q=Welco…
And yet, the assumption is false. This is not just my opinion, it's been demonstrated and accepted as uncontroversial for decades
In mathematics, for example, most knowledge exists in the form of "knowhow", and there are already large-scale efforts underway to capture it x.com/SAIRfoundation…
So what's going on? In this article, I explain x.com/Dr_JohnFletche…
@sequoia @RichardSSutton
Rohan Paul@rohanpaul_ai
Google DeepMind veteran David Silver just launched a London AI lab Ineffable Intelligence, and raised $1B at a $4B valuation, bets on radically new type of Reinforcement Learning to build superintelligence. Silver’s core argument is that large language models — the architecture behind ChatGPT, Claude, Gemini and every major AI system in commercial use today — are fundamentally limited. They learn from human-generated data. They can synthesise, summarise and extend what humans have already written or thought. But they cannot, in Silver’s view, discover genuinely new knowledge. Ineffable Intelligence aims to build what Silver has described as “an endlessly learning superintelligence that self-discovers the foundations of all knowledge.” The approach is rooted in reinforcement learning — the branch of AI Silver has spent his entire career advancing. --- the-decoder. com/deepmind-veteran-david-silver-raises-1b-seed-round-to-build-superintelligence-without-llms/
English
My retweetledi
I bought more $TIG after reading John's post the other day. It was fantastic. But I want to take a second to highlight his co-founder Philip David.
The guy who built ARM's licensing strategy now runs IP at $TIG.
Philip was General Council at ARM for 15 years. SVP of IP during the SoftBank acquisition. He designed the framework that lets ARM collect royalties on 95% of the world's smartphones without manufacturing a single chip. Yeah. Read that again.
ARM doesn't make anything. They create the IP. They license it. They get paid every time someone ships.
That's exactly what $TIG is doing for algorithms.
Open Data License for researchers. Commercial License for enterprises. Every algorithm stays open, but commercial use generates revenue that flows back to contributors.
The man who turned processor IP into a $150B+ licensing machine is now applying the same playbook to computational methods.
Most people in crypto have never shipped a real licensing framework.
$TIG's co-founder spent a decade and a half building the most successful one in semiconductor history.
Just some food for thought.
John Fletcher (𝔦, 𝔦)@Dr_JohnFletcher
Andrej, I’m John Fletcher. I have a PhD in mathematics and theoretical physics from Cambridge, and since 2016 I have been working full-time on the problem of how to coordinate untrusted distributed compute for algorithmic innovation. I listened to your No Priors conversation and recognised the architecture you were describing: commits that build on each other, computational asymmetry (hard to find, cheap to verify), an untrusted pool of workers collaborating through a blockchain-like structure. The result is The Innovation Game (TIG), which has been in continuous operation since mid-2024. The correspondence is so close that I thought it worth writing. The short version: roughly 7,000 Benchmarkers test algorithms submitted by Innovators by solving instances of asymmetric computational challenges (SAT, Vehicle Routing, Quadratic Knapsack, Vector Search, among others). This testing is "proof of work" in the technical sense of Dwork and Naor (1992). Innovators earn rewards proportional to adoption by the Benchmarkers. The repository of algorithms is open source (github.com/tig-foundation…). The system is already producing state-of-the-art results. For the Quadratic Knapsack Problem, 476 iterative submissions by independent contributors brought solution quality to a level that now exceeds methods published by Hochbaum et al. in the European Journal of Operational Research (2025). We are working with Thibaut Vidal (Polytechnique Montréal), who has submitted a state-of-the-art vehicle routing algorithm directly to TIG, and with Yuji Nakatsukasa (Oxford) and Dario Paccagnan (Imperial College London), among many others. One of TIG’s active challenges is directly relevant to your autoresearch work: an optimiser for neural network training (play.tig.foundation/challenges?cha…), where Innovators compete to develop an improved optimiser (see screenshot). One way in which TIG extends the vision is on the economic side. In our view, a monetary incentive is required, otherwise the open strand simply cannot compete at scale. TIG’s open source dual licensing model (designed by my co-founder Philip David, who was General Counsel at Arm Holdings for over a decade, and was the artchitect of ARMs licensing strategy) is intended to solve that problem. I expect we have each thought about parts of this that the other hasn’t. Happy to talk whenever suits. John Fletcher tig.foundation
English
My retweetledi
@karpathy Hi Andrej, Please feel free to demolish our argument (with the help of AI or otherwise!) x.com/Dr_JohnFletche…
John Fletcher (𝔦, 𝔦)@Dr_JohnFletcher
Andrej, I’m John Fletcher. I have a PhD in mathematics and theoretical physics from Cambridge, and since 2016 I have been working full-time on the problem of how to coordinate untrusted distributed compute for algorithmic innovation. I listened to your No Priors conversation and recognised the architecture you were describing: commits that build on each other, computational asymmetry (hard to find, cheap to verify), an untrusted pool of workers collaborating through a blockchain-like structure. The result is The Innovation Game (TIG), which has been in continuous operation since mid-2024. The correspondence is so close that I thought it worth writing. The short version: roughly 7,000 Benchmarkers test algorithms submitted by Innovators by solving instances of asymmetric computational challenges (SAT, Vehicle Routing, Quadratic Knapsack, Vector Search, among others). This testing is "proof of work" in the technical sense of Dwork and Naor (1992). Innovators earn rewards proportional to adoption by the Benchmarkers. The repository of algorithms is open source (github.com/tig-foundation…). The system is already producing state-of-the-art results. For the Quadratic Knapsack Problem, 476 iterative submissions by independent contributors brought solution quality to a level that now exceeds methods published by Hochbaum et al. in the European Journal of Operational Research (2025). We are working with Thibaut Vidal (Polytechnique Montréal), who has submitted a state-of-the-art vehicle routing algorithm directly to TIG, and with Yuji Nakatsukasa (Oxford) and Dario Paccagnan (Imperial College London), among many others. One of TIG’s active challenges is directly relevant to your autoresearch work: an optimiser for neural network training (play.tig.foundation/challenges?cha…), where Innovators compete to develop an improved optimiser (see screenshot). One way in which TIG extends the vision is on the economic side. In our view, a monetary incentive is required, otherwise the open strand simply cannot compete at scale. TIG’s open source dual licensing model (designed by my co-founder Philip David, who was General Counsel at Arm Holdings for over a decade, and was the artchitect of ARMs licensing strategy) is intended to solve that problem. I expect we have each thought about parts of this that the other hasn’t. Happy to talk whenever suits. John Fletcher tig.foundation
English
My retweetledi
If you’ve ever wondered what kinds of gains @tigfoundation’s various AI-oriented challenges could unlock, this should give you a rough idea. Google’s new TurboQuant algo reduces memory usage by 6x while boosting speed 8x. It does this all without compromising on accuracy.
With $TIG’s Optimizer challenge in particular, the performance gains for LLMs would be unimaginably higher. $TIG’s prior algorithmic breakthroughs demonstrate that it’s only a matter of time.
Google Research@GoogleResearch
Introducing TurboQuant: Our new compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x speedup, all with zero accuracy loss, redefining AI efficiency. Read the blog to learn how it achieves these results: goo.gle/4bsq2qI
English
My retweetledi
This week, Andrej Karpathy independently described TIG without being aware of its existence.
@dr_johnfletcher’s reply quickly became- by orders of magnitude - our most viral post ever.
Today at 5PM GMT, join the good Doctor and @0x_Asuka as they discuss what this means for TIG and science as a whole.
English
My retweetledi
@karpathy @Dr_JohnFletcher has been working on exactly what you’re describing for 10+ years. Since 2024, it has been live, continuously refined, and even produced state-of-the-art algorithms that far exceed the those that came before them in their respective domains. @tigfoundation is genuinely one of the most remarkable open source initiatives of our era and definitely worth a closer look!
English
My retweetledi
John Fletcher just put $TIG in front of Karpathy.
Karpathy recently said (paraphrased) on No Priors: "My designs that incorporate an untrusted pool of workers actually look a little bit more like a blockchain. Instead of blocks, you have commits, and these commits can build on each other."
Yeah. That's @tigfoundation.
@karpathy independently described the $TIG architecture without knowing it exists.
$8.5M mc.
The asymmetry here is absurd.
Most crypto proof of work is computational theater. Trillions in energy burning SHA-256 hashes that produce nothing.
$TIG is literally proof of useful work. The protocol splits into two roles: Benchmarkers compete on compute, Innovators compete on writing better algorithms. The best algorithm wins. Every round. Automatically.
Imo, this is one of the most undervalued mechanism designs in crypto. Upside here is silly to me.
John Fletcher (𝔦, 𝔦)@Dr_JohnFletcher
Andrej, I’m John Fletcher. I have a PhD in mathematics and theoretical physics from Cambridge, and since 2016 I have been working full-time on the problem of how to coordinate untrusted distributed compute for algorithmic innovation. I listened to your No Priors conversation and recognised the architecture you were describing: commits that build on each other, computational asymmetry (hard to find, cheap to verify), an untrusted pool of workers collaborating through a blockchain-like structure. The result is The Innovation Game (TIG), which has been in continuous operation since mid-2024. The correspondence is so close that I thought it worth writing. The short version: roughly 7,000 Benchmarkers test algorithms submitted by Innovators by solving instances of asymmetric computational challenges (SAT, Vehicle Routing, Quadratic Knapsack, Vector Search, among others). This testing is "proof of work" in the technical sense of Dwork and Naor (1992). Innovators earn rewards proportional to adoption by the Benchmarkers. The repository of algorithms is open source (github.com/tig-foundation…). The system is already producing state-of-the-art results. For the Quadratic Knapsack Problem, 476 iterative submissions by independent contributors brought solution quality to a level that now exceeds methods published by Hochbaum et al. in the European Journal of Operational Research (2025). We are working with Thibaut Vidal (Polytechnique Montréal), who has submitted a state-of-the-art vehicle routing algorithm directly to TIG, and with Yuji Nakatsukasa (Oxford) and Dario Paccagnan (Imperial College London), among many others. One of TIG’s active challenges is directly relevant to your autoresearch work: an optimiser for neural network training (play.tig.foundation/challenges?cha…), where Innovators compete to develop an improved optimiser (see screenshot). One way in which TIG extends the vision is on the economic side. In our view, a monetary incentive is required, otherwise the open strand simply cannot compete at scale. TIG’s open source dual licensing model (designed by my co-founder Philip David, who was General Counsel at Arm Holdings for over a decade, and was the artchitect of ARMs licensing strategy) is intended to solve that problem. I expect we have each thought about parts of this that the other hasn’t. Happy to talk whenever suits. John Fletcher tig.foundation
English
My retweetledi
Here’s definitive proof that @tigfoundation is way ahead of the curve. Open AI cofounder and former head of AI @ Tesla @karpathy goes on the record to describe how a decentralized system for open-source algorithm development is exactly what we need to propel AI forward. He even quips that it would look ‘kind of like a blockchain’. This is exactly what @Dr_JohnFletcher and co. have been working on for 10+ years. It’s already online and proven to work- most people just don’t know about it yet.
This is why I’ve been glazing @tigfoundation for years. Every week, more and more CEOs/thought leaders are admitting that what they’ve built will usher in an entirely new paradigm for science and AI.
$TIG is inevitable.
John Fletcher (𝔦, 𝔦)@Dr_JohnFletcher
Andrej, I’m John Fletcher. I have a PhD in mathematics and theoretical physics from Cambridge, and since 2016 I have been working full-time on the problem of how to coordinate untrusted distributed compute for algorithmic innovation. I listened to your No Priors conversation and recognised the architecture you were describing: commits that build on each other, computational asymmetry (hard to find, cheap to verify), an untrusted pool of workers collaborating through a blockchain-like structure. The result is The Innovation Game (TIG), which has been in continuous operation since mid-2024. The correspondence is so close that I thought it worth writing. The short version: roughly 7,000 Benchmarkers test algorithms submitted by Innovators by solving instances of asymmetric computational challenges (SAT, Vehicle Routing, Quadratic Knapsack, Vector Search, among others). This testing is "proof of work" in the technical sense of Dwork and Naor (1992). Innovators earn rewards proportional to adoption by the Benchmarkers. The repository of algorithms is open source (github.com/tig-foundation…). The system is already producing state-of-the-art results. For the Quadratic Knapsack Problem, 476 iterative submissions by independent contributors brought solution quality to a level that now exceeds methods published by Hochbaum et al. in the European Journal of Operational Research (2025). We are working with Thibaut Vidal (Polytechnique Montréal), who has submitted a state-of-the-art vehicle routing algorithm directly to TIG, and with Yuji Nakatsukasa (Oxford) and Dario Paccagnan (Imperial College London), among many others. One of TIG’s active challenges is directly relevant to your autoresearch work: an optimiser for neural network training (play.tig.foundation/challenges?cha…), where Innovators compete to develop an improved optimiser (see screenshot). One way in which TIG extends the vision is on the economic side. In our view, a monetary incentive is required, otherwise the open strand simply cannot compete at scale. TIG’s open source dual licensing model (designed by my co-founder Philip David, who was General Counsel at Arm Holdings for over a decade, and was the artchitect of ARMs licensing strategy) is intended to solve that problem. I expect we have each thought about parts of this that the other hasn’t. Happy to talk whenever suits. John Fletcher tig.foundation
English
John Fletcher (𝔦, 𝔦)@Dr_JohnFletcher
Andrej, I’m John Fletcher. I have a PhD in mathematics and theoretical physics from Cambridge, and since 2016 I have been working full-time on the problem of how to coordinate untrusted distributed compute for algorithmic innovation. I listened to your No Priors conversation and recognised the architecture you were describing: commits that build on each other, computational asymmetry (hard to find, cheap to verify), an untrusted pool of workers collaborating through a blockchain-like structure. The result is The Innovation Game (TIG), which has been in continuous operation since mid-2024. The correspondence is so close that I thought it worth writing. The short version: roughly 7,000 Benchmarkers test algorithms submitted by Innovators by solving instances of asymmetric computational challenges (SAT, Vehicle Routing, Quadratic Knapsack, Vector Search, among others). This testing is "proof of work" in the technical sense of Dwork and Naor (1992). Innovators earn rewards proportional to adoption by the Benchmarkers. The repository of algorithms is open source (github.com/tig-foundation…). The system is already producing state-of-the-art results. For the Quadratic Knapsack Problem, 476 iterative submissions by independent contributors brought solution quality to a level that now exceeds methods published by Hochbaum et al. in the European Journal of Operational Research (2025). We are working with Thibaut Vidal (Polytechnique Montréal), who has submitted a state-of-the-art vehicle routing algorithm directly to TIG, and with Yuji Nakatsukasa (Oxford) and Dario Paccagnan (Imperial College London), among many others. One of TIG’s active challenges is directly relevant to your autoresearch work: an optimiser for neural network training (play.tig.foundation/challenges?cha…), where Innovators compete to develop an improved optimiser (see screenshot). One way in which TIG extends the vision is on the economic side. In our view, a monetary incentive is required, otherwise the open strand simply cannot compete at scale. TIG’s open source dual licensing model (designed by my co-founder Philip David, who was General Counsel at Arm Holdings for over a decade, and was the artchitect of ARMs licensing strategy) is intended to solve that problem. I expect we have each thought about parts of this that the other hasn’t. Happy to talk whenever suits. John Fletcher tig.foundation
QME
@karpathy We’re translating our thoughts into specs and then LLMs push them into code. We also have store some artifacts in English in .md files.
When do you think we fully go to spec driven development to not having to do this translation? And if not, what’s the end state of code?
English
Thank you Sarah, my pleasure to come on the pod! And happy to do some more Q&A in the replies.
sarah guo@saranormous
Caught up with @karpathy for a new @NoPriorsPod: on the phase shift in engineering, AI psychosis, claws, AutoResearch, the opportunity for a SETI-at-Home like movement in AI, the model landscape, and second order effects 02:55 - What Capability Limits Remain? 06:15 - What Mastery of Coding Agents Looks Like 11:16 - Second Order Effects of Coding Agents 15:51 - Why AutoResearch 22:45 - Relevant Skills in the AI Era 28:25 - Model Speciation 32:30 - Collaboration Surfaces for Humans and AI 37:28 - Analysis of Jobs Market Data 48:25 - Open vs. Closed Source Models 53:51 - Autonomous Robotics and Atoms 1:00:59 - MicroGPT and Agentic Education 1:05:40 - End Thoughts
English
John Fletcher (𝔦, 𝔦)@Dr_JohnFletcher
Andrej, I’m John Fletcher. I have a PhD in mathematics and theoretical physics from Cambridge, and since 2016 I have been working full-time on the problem of how to coordinate untrusted distributed compute for algorithmic innovation. I listened to your No Priors conversation and recognised the architecture you were describing: commits that build on each other, computational asymmetry (hard to find, cheap to verify), an untrusted pool of workers collaborating through a blockchain-like structure. The result is The Innovation Game (TIG), which has been in continuous operation since mid-2024. The correspondence is so close that I thought it worth writing. The short version: roughly 7,000 Benchmarkers test algorithms submitted by Innovators by solving instances of asymmetric computational challenges (SAT, Vehicle Routing, Quadratic Knapsack, Vector Search, among others). This testing is "proof of work" in the technical sense of Dwork and Naor (1992). Innovators earn rewards proportional to adoption by the Benchmarkers. The repository of algorithms is open source (github.com/tig-foundation…). The system is already producing state-of-the-art results. For the Quadratic Knapsack Problem, 476 iterative submissions by independent contributors brought solution quality to a level that now exceeds methods published by Hochbaum et al. in the European Journal of Operational Research (2025). We are working with Thibaut Vidal (Polytechnique Montréal), who has submitted a state-of-the-art vehicle routing algorithm directly to TIG, and with Yuji Nakatsukasa (Oxford) and Dario Paccagnan (Imperial College London), among many others. One of TIG’s active challenges is directly relevant to your autoresearch work: an optimiser for neural network training (play.tig.foundation/challenges?cha…), where Innovators compete to develop an improved optimiser (see screenshot). One way in which TIG extends the vision is on the economic side. In our view, a monetary incentive is required, otherwise the open strand simply cannot compete at scale. TIG’s open source dual licensing model (designed by my co-founder Philip David, who was General Counsel at Arm Holdings for over a decade, and was the artchitect of ARMs licensing strategy) is intended to solve that problem. I expect we have each thought about parts of this that the other hasn’t. Happy to talk whenever suits. John Fletcher tig.foundation
QME
Great questions! Starting backwards with (3), I'd hope AIs can feel like Rocky from Project Hail Mary (it's top of mind having seen it yesterday), like a partner and a teammate. As one small example that stuck with me recently, when Claude found the Sonos system on my LAN, it could have said something like
"Successfully found the sonos server..."
Instead it said something like
"We're in!..."
Small example, but I feel like there's a sense that we're trying to achieve something together, etc. Possibly others have their own versions. I still think it could be better, for example Rocky has his own distinct personality, opinions, quirks, I don't feel like AIs have these aspects, they feel a little too safe, and overall still a little bit too corporate.
(2) No, the field is not being intentional enough, which ties to (1), I don't think personality requires new technology, I think it looks more like long SOUL.md files, possibly distilled into weights, and an organizational choice to install someone who really cares about this dimension and who drives it.
English
John Fletcher (𝔦, 𝔦)@Dr_JohnFletcher
Andrej, I’m John Fletcher. I have a PhD in mathematics and theoretical physics from Cambridge, and since 2016 I have been working full-time on the problem of how to coordinate untrusted distributed compute for algorithmic innovation. I listened to your No Priors conversation and recognised the architecture you were describing: commits that build on each other, computational asymmetry (hard to find, cheap to verify), an untrusted pool of workers collaborating through a blockchain-like structure. The result is The Innovation Game (TIG), which has been in continuous operation since mid-2024. The correspondence is so close that I thought it worth writing. The short version: roughly 7,000 Benchmarkers test algorithms submitted by Innovators by solving instances of asymmetric computational challenges (SAT, Vehicle Routing, Quadratic Knapsack, Vector Search, among others). This testing is "proof of work" in the technical sense of Dwork and Naor (1992). Innovators earn rewards proportional to adoption by the Benchmarkers. The repository of algorithms is open source (github.com/tig-foundation…). The system is already producing state-of-the-art results. For the Quadratic Knapsack Problem, 476 iterative submissions by independent contributors brought solution quality to a level that now exceeds methods published by Hochbaum et al. in the European Journal of Operational Research (2025). We are working with Thibaut Vidal (Polytechnique Montréal), who has submitted a state-of-the-art vehicle routing algorithm directly to TIG, and with Yuji Nakatsukasa (Oxford) and Dario Paccagnan (Imperial College London), among many others. One of TIG’s active challenges is directly relevant to your autoresearch work: an optimiser for neural network training (play.tig.foundation/challenges?cha…), where Innovators compete to develop an improved optimiser (see screenshot). One way in which TIG extends the vision is on the economic side. In our view, a monetary incentive is required, otherwise the open strand simply cannot compete at scale. TIG’s open source dual licensing model (designed by my co-founder Philip David, who was General Counsel at Arm Holdings for over a decade, and was the artchitect of ARMs licensing strategy) is intended to solve that problem. I expect we have each thought about parts of this that the other hasn’t. Happy to talk whenever suits. John Fletcher tig.foundation
QME
Sure! Andrej Karpathy (top AI expert) is saying:
AI "agents" (smart tools that auto-write computer code for you) often make the code messy and hard to read. They ignore his rules like "keep every line super simple, one job only."
Instead, they cram too much into one line, copy-paste junk, and bloat things up.
He used to fight it but now shrugs— it works, even if it's not pretty.
He also notes using AI to "judge" good code has limits long-term, but we're not there yet.
Basically, AI coders are powerful but still need human taste for clean work.
English
John Fletcher (𝔦, 𝔦)@Dr_JohnFletcher
Andrej, I’m John Fletcher. I have a PhD in mathematics and theoretical physics from Cambridge, and since 2016 I have been working full-time on the problem of how to coordinate untrusted distributed compute for algorithmic innovation. I listened to your No Priors conversation and recognised the architecture you were describing: commits that build on each other, computational asymmetry (hard to find, cheap to verify), an untrusted pool of workers collaborating through a blockchain-like structure. The result is The Innovation Game (TIG), which has been in continuous operation since mid-2024. The correspondence is so close that I thought it worth writing. The short version: roughly 7,000 Benchmarkers test algorithms submitted by Innovators by solving instances of asymmetric computational challenges (SAT, Vehicle Routing, Quadratic Knapsack, Vector Search, among others). This testing is "proof of work" in the technical sense of Dwork and Naor (1992). Innovators earn rewards proportional to adoption by the Benchmarkers. The repository of algorithms is open source (github.com/tig-foundation…). The system is already producing state-of-the-art results. For the Quadratic Knapsack Problem, 476 iterative submissions by independent contributors brought solution quality to a level that now exceeds methods published by Hochbaum et al. in the European Journal of Operational Research (2025). We are working with Thibaut Vidal (Polytechnique Montréal), who has submitted a state-of-the-art vehicle routing algorithm directly to TIG, and with Yuji Nakatsukasa (Oxford) and Dario Paccagnan (Imperial College London), among many others. One of TIG’s active challenges is directly relevant to your autoresearch work: an optimiser for neural network training (play.tig.foundation/challenges?cha…), where Innovators compete to develop an improved optimiser (see screenshot). One way in which TIG extends the vision is on the economic side. In our view, a monetary incentive is required, otherwise the open strand simply cannot compete at scale. TIG’s open source dual licensing model (designed by my co-founder Philip David, who was General Counsel at Arm Holdings for over a decade, and was the artchitect of ARMs licensing strategy) is intended to solve that problem. I expect we have each thought about parts of this that the other hasn’t. Happy to talk whenever suits. John Fletcher tig.foundation
QME
Caught up with @karpathy for a new @NoPriorsPod: on the phase shift in engineering, AI psychosis, claws, AutoResearch, the opportunity for a SETI-at-Home like movement in AI, the model landscape, and second order effects
02:55 - What Capability Limits Remain?
06:15 - What Mastery of Coding Agents Looks Like
11:16 - Second Order Effects of Coding Agents
15:51 - Why AutoResearch
22:45 - Relevant Skills in the AI Era
28:25 - Model Speciation
32:30 - Collaboration Surfaces for Humans and AI
37:28 - Analysis of Jobs Market Data
48:25 - Open vs. Closed Source Models
53:51 - Autonomous Robotics and Atoms
1:00:59 - MicroGPT and Agentic Education
1:05:40 - End Thoughts
English