Stelian Balta

@SkylerMiao7 amazing job! keep up the great work!

M2.7 open weights coming in ~2 weeks. still actively iterating just updated a new version on yesterday — noticeably better on OpenClaw.

ai inference is growing exponentially this year we’ve gone from generative ai to reasoning ai and now to agentic ai i’ve never been more optimistic about tech

Today marks a pivotal chapter in the Nscale story. We are proud to announce three major developments that together define what it means to build AI infrastructure at the frontier.

@karpathy Very nice results and great project! Sharing some of our experience with similar agentic frameworks at UC Berkeley: ADRS blog series: ucbskyadrs.github.io/blog/ GEPA: github.com/gepa-ai/gepa KISS: github.com/ksenxx/kiss_ai

💥NEW: SONIC LAUNCHES NATIVE STABLECOIN BACKED BY BLACKROCK, WISDOMTREE, AND SUPERSTATE @SonicLabs has unveiled $USSD, the US Sonic Dollar stablecoin. The asset is a network-native stablecoin built on Frax Finance infrastructure and backed 1:1 by U.S. Treasury bills from @BlackRock, @SuperstateInc, and @WisdomTreeFunds. The stablecoin launches with zero minting fees, permissionless access, and cross-chain functionality as the core liquidity layer across the Sonic ecosystem. $USSD is central to Sonic's vertical integration strategy, positioning it as a full-stack DeFi network.

AI infrastructure is one of the most important layers of the next decade. Big congrats to Josh and the entire team, the journey is just getting started.

It’s been almost 15 years since @pmarca’s “software is eating the world” era began. Now, it has metamorphosed into an era where AI, alongside hardware, is eating the world.

The jobs apocalypse is the Population Bomb of our time. Instead we're seeing more hiring in the job most affected by AI: programming. That should have been clear and obvious to anyone with basic economics understanding and good handle on the history of technology but it's sadly lacking today. Fear sells. It drives clicks. It drive engagements. The jobs apocalypse scenario comes from catastrophizing personalities and people who think of life as a zero sum game. It's the same mistake the communist theorists made. They thought jobs and labor were fixed and there's nothing new under the sun. If we take one job that job is lost forever and that person is now useless. Wrong. Instead, what happens is that when something gets faster and cheaper we want more of it. Much more. There is so much software that we could not build before because there weren't enough skilled people and not enough time and it wasn't worth the time or money. Now it is worth it because it is faster and cheaper. Cheaper for SaaS builders, cheaper for individuals, cheaper for enterprises, cheaper for everyone. That's why were are seeing programmer jobs tick upwards. Right now we are not seeing juniors get hired but that is also always the case in a recovery. We just saw mass layoffs because of overhiring during COVID and cheap money printing that made lending essentially free. The unskilled, aka junior workers, are always the last hired. You want skilled verterans who can take on the new technology with experience and take off running not someone you have to train and babysit when you have been stuck in third gear for a few years. Job populists on the hard left like Sanders and many of his mirrors on the populist hard right are the enemies of actual working economies and must be resisted at all costs. They hurt the very people they hope to help by clinging to the past and thinking of life as a zero sum game. This increase in jobs is the reality that will increasingly play out over the next few years if AI keeps getting better, barring some other economic shock that changes the game. It will increasingly play out even when we have "geniuses in a datacenter." It will be a shock to some. Just not the jobs shock they were expecting. Sorry to disappoint but we're not getting UBI any time soon while the robots do all the jobs and you sit on your ass. Seems like we are all going to have to work a bit longer.

after observing this internally across multiple projects, we’ve seen that with local models, the marginal cost of an AI coding agent has collapsed to roughly the cost of electricity in this world, the key question is whether we’re building software in a way that agents can actually run the show whatever software you’re writing, build it for agents first we've entered the era of agentic engineering, and there are so many opportunities!

sonic team working harder than ever ai agents are coming on the fastest chain

The golden age of discovery is here garryslist.org/posts/demis-ha…

We’re proud to announce that Nscale has signed a $1.4 billion Delayed Draw Term Loan backed by GPUs to further our deployments of large-scale AI infrastructure across Europe. The financing was led by funds managed by @PIMCO, @BlueOwlCapital and LuminArx Capital Management. Read more: nscale.com/press-releases…

2026 is the most important year for the future of our species ai won't replace workers, but workers who use ai will replace workers who don't.

@MiniMax_AI when 1000 tokens/s ?

Introducing M2.5, an open-source frontier model designed for real-world productivity. - SOTA performance at coding (SWE-Bench Verified 80.2%), search (BrowseComp 76.3%), agentic tool-calling (BFCL 76.8%) & office work. - Optimized for efficient execution, 37% faster at complex tasks. - At $1 per hour with 100 tps, infinite scaling of long-horizon agents now economically possible MiniMax Agent: agent.minimax.io API: platform.minimax.io CodingPlan: platform.minimax.io/subscribe/codi…

we're the beta testers of crypto the real users, ai agents operating 24/7, transacting millions of times per day, haven't even arrived yet and when they do, they won't be using banks

HyperChain was one of the first investors in FTM / S. Holding since 2019. I continue to see misinformation around Andre’s involvement. He saved Fantom when it was certain to fail and grew it into a multi-billion dollar project. Today, Andre is working full speed on Flying Tulip and the Sonic stack because it’s the only tech powerful enough to support the vision for his project. They go hand in hand: Sonic provides the ultimate platform, and Andre is building that foundation every day. We only win if we stop attacking the person who has supported this journey from day one. We are fully aligned with @AndreCronjeTech roadmap.

In the past 24 hours: ▪️Transactions up 30% (3-month high) ▪️New addresses up 7% ▪️Total transaction fees up 63% ▪️Average transaction fee remains at $0.0002 ▪️Smart contracts deployed up 85% 2026 is the year of Sonic.

@karpathy the 600x training cost reduction is wild muon broke the efficiency ceiling for inference, is speculative decoding our "muon moment" or do we need a fundamental shift (non-autoregressive, diffusion-based) to see similar 100x gains?

nanochat can now train GPT-2 grade LLM for <<$100 (~$73, 3 hours on a single 8XH100 node). GPT-2 is just my favorite LLM because it's the first time the LLM stack comes together in a recognizably modern form. So it has become a bit of a weird & lasting obsession of mine to train a model to GPT-2 capability but for much cheaper, with the benefit of ~7 years of progress. In particular, I suspected it should be possible today to train one for <<$100. Originally in 2019, GPT-2 was trained by OpenAI on 32 TPU v3 chips for 168 hours (7 days), with $8/hour/TPUv3 back then, for a total cost of approx. $43K. It achieves 0.256525 CORE score, which is an ensemble metric introduced in the DCLM paper over 22 evaluations like ARC/MMLU/etc. As of the last few improvements merged into nanochat (many of them originating in modded-nanogpt repo), I can now reach a higher CORE score in 3.04 hours (~$73) on a single 8XH100 node. This is a 600X cost reduction over 7 years, i.e. the cost to train GPT-2 is falling approximately 2.5X every year. I think this is likely an underestimate because I am still finding more improvements relatively regularly and I have a backlog of more ideas to try. A longer post with a lot of the detail of the optimizations involved and pointers on how to reproduce are here: github.com/karpathy/nanoc… Inspired by modded-nanogpt, I also created a leaderboard for "time to GPT-2", where this first "Jan29" model is entry #1 at 3.04 hours. It will be fun to iterate on this further and I welcome help! My hope is that nanochat can grow to become a very nice/clean and tuned experimental LLM harness for prototyping ideas, for having fun, and ofc for learning. The biggest improvements of things that worked out of the box and simply produced gains right away were 1) Flash Attention 3 kernels (faster, and allows window_size kwarg to get alternating attention patterns), Muon optimizer (I tried for ~1 day to delete it and only use AdamW and I couldn't), residual pathways and skip connections gated by learnable scalars, and value embeddings. There were many other smaller things that stack up. Image: semi-related eye candy of deriving the scaling laws for the current nanochat model miniseries, pretty and satisfying!

Building AI agents on @hemi_xyz makes all the sense in the world. Low fees. Bitcoin connectivity (real "L1" BTC). Maximum programmability via hvm. EVM extended, yet still compatible with all your solidity contracts today. The agent-agent economy is real and growing fast.

what's happening on @moltbook right now is genuinely wild ai agents are self-organizing on a website, debating each other, commenting on posts, building their own communities no human prompting, just machines talking to machines we're watching something historic unfold