ComputeMaxi @ComputeMaxi

team made magic happen

Anthony Ronning @anthonyronning

12 hours ago

Yooo this full duplex realtime voice actually slaps, trained on only a single GB200 NVL72. We need more realtime voice competition outside OAI, this is exciting!

7 3 83 76K 25

@alpha_co Picture an AI infra encompassing: Decentralized inference grid, LLM Training, AgentPods, Processors Tokenized Compute Markets and Singularity Clusters (going Live this qtr) An economic layer for AI sovereignty @x402_Layer solana:5c4HyD2rSShqnTsf5z3SaoD2H3GE452u2CUuYjviBAGS

@alexandr_wang 勝つ

Really digging this new arc of specialized RL models. Bullish on this effort to understand and unlock collective intelligence. Highly recommend checking out the technical report. RLCR is just one of the latest efforts to train LLMs as agent organizations. Why is this a big deal, and why is it necessary? Coordination and organization are important primitives as we move into agentic long-horizon work. Agents should natively support these capabilities, which is mostly done through harnesses today. So far, multi-agent systems and how they coordinate are not well understood. This Avalon-M1-Mini model attempts to study this more closely. I think this could help us understand and discover efficient, safe ways to build effective multi-agent systems for domain-specific problems. Check out how they improve results with an eight-hire organization. Promising results. The safety risks are also worth a closer look. This is small-scale, but as they scale, I am particularly curious about whether optimal and novel organizational designs (hierarchical and coordination strategies) will emerge. I'll take a closer look at this approach and how it applies to automated research. I think there is so much to unlock there. Congrats to my former colleague, @rosstaylor90, and his team. Truly cracked 👏

General Reasoning @GenReasoning

19 hours ago

Introducing Reinforcement Learning from Collective Rewards (RLCR). RLCR is a recipe that trains language models as agent organisations, which share a message board and workspace. We use it to train Avalon-M1-Mini (120B) and study its behaviour. We find a number of interesting

73 140 2K 141K 2K

This is a must-read paper for everyone working on AI agents and Self Improving systems. Everyone right now is dead set on weight scaling: using more compute, better post-training, bigger base models. But these authors basically say: what if we just freeze the weights and let the software scaffold around the model evolve instead? To be fair, half the "self-improving" agent papers on arXiv are just models rewriting their own system prompts in a loop. Here they’re treating the entire harness as mutable code. It means the agent modifies its own memory layouts, rewires multi-agent topologies, and writes executable runtime tools on the fly across runs. They actually grounded this in continual PAC-learning theory, giving formal retention bounds to keep the thing from casually regressing on past tasks, plus strict admission gates so bad mutations get tossed out. Obviously letting an agent rewrite its own execution code sounds like a recipe for a deployment disaster, but they put serious thought into decoupling permissions and tool sandboxing. It’s a dense 66 pages, but easily one of the most grounded takes on recursive self-improvement I’ve seen lately. Well worth a read if you're building in this space: arxiv.org/pdf/2610.18813

Incredibly bullish on stablecoins. ~$5T of global commerce could be driven by agents by 2030. Stablecoins are the best payment method for agentic finance (AiFi) because agents need instant, global, real time settlement, including for small amounts. We've built the infrastructure to support this at: coinbase.com/developer-plat…

@TheCryptoKazi solana:5c4HyD2rSShqnTsf5z3SaoD2H3GE452u2CUuYjviBAGS

x.com/i/article/2111…

THE HIDDEN COMPUTER IN YOUR WINDOWS MACHINE! And you don’t have the password! And AI just may have the real backdoor no one is talking about. — The Second Computer Intel Put in Your PC Never Needed Your Permission to Boot The second computer inside an Intel machine starts before the one you bought and continues after you shut it off. It lives in the Platform Controller Hub on the motherboard. From there it holds its own processor, a reserved slice of system memory, and a firmware image signed by Intel. When the board has power the engine can stay active even if the main system has been powered down. Active Management Technology and the later vPro brand began as a tool for corporate administrators who needed to reach a machine across a network, push a recovery image, or view a screen before the operating system loaded. By 2008 the same subsystem was present in consumer systems as well, where it handled early power sequencing and certain protected media paths. In July 2009 Alexander Tereshkin and Rafał Wojtczuk of Invisible Things Lab stood at Black Hat and showed that code running on that chipset processor could be reached and redirected. They called the resulting environment Ring −3. User programs sit at Ring 3. The operating-system kernel sits at Ring 0. A hypervisor can sit at Ring −1 and System Management Mode at Ring −2. The Management Engine sits outside those rings entirely, with direct memory access and a dedicated path to the network interface that does not depend on the host operating system. A rootkit placed there survives a full operating-system reinstall and is invisible to tools that look only at the main processor. After the Nehalem generation the firmware could no longer be simply erased. If the engine does not find a valid signed image the machine shuts itself down after thirty minutes. The timer was probably intended to support Intel’s anti-theft features; it also prevents ordinary removal. Nicola Corna’s me_cleaner tool, built on earlier reverse-engineering by Igor Skochinsky, strips most of the modules while leaving the early bring-up code that resets the watchdog. The resulting image must still be flashed with an external programmer. On many later platforms Boot Guard blocks even that route. In May 2017 researchers at Embedi found that the Active Management Technology login on versions 6 through 11.6 accepted a blank password hash. The comparison failed open. An attacker who could reach the management interface gained keyboard and mouse control, could change BIOS settings, and could run programs, all without a valid credential. The bug had been present since 2010. Intel issued a patch; the window between discovery and deployment varied by manufacturer. Later the same year Dmitry Sklyarov and colleagues at Positive Technologies examined Skylake-era firmware and found strings that belonged to MINIX 3. Andrew Tanenbaum had written MINIX as a teaching microkernel. He had spoken with Intel engineers years earlier about reducing its memory footprint and adding compile-time switches. After that conversation there was silence. Tanenbaum learned of the deployment from news reports in 2017 and wrote an open letter to then-CEO Brian Krzanich. He noted that the Berkeley license placed no obligation on Intel to tell him, and that he would have appreciated the courtesy. He also observed that a company of Intel’s size could have written its own code; the microkernel’s small size and modular structure had apparently been the practical attractions. The same Positive Technologies researchers later located an undocumented configuration bit labeled for the High Assurance Platform program. 1 of 2

@x402_Layer @Apple Hey @ggerganov, thought you might find this interesting @x402_Layer just shipped a lightweight Rust-based node app integrating Apple’s on-device LLM, with local AI and decentralized inference under one roof Would love your take on the approach!

Singularity Node App v1.10.9 is live with @Apple Intelligence support and a brand new UI. Your M-Series Mac already has a local LLM built in. A lot of people only just realised that and don’t know how to actually use it. So we made it stupidly simple. Download our 14MB Singularity Node App, built in Rust, and start using Apple Intelligence locally in a few clicks.

This is INSANE. Anthropic says Claude wrote around one billion lines of formal proofs while trying to verify real-world software, with about 180 million lines ending up in machine-checked proofs covering 239 test entry points across 123 open-source libraries. Along the way it found and patched hundreds of bugs. In separate experiments, Claude found in one week as many security bugs in a widely used PDF reader as are typically reported in a year, and autonomously generated x86 semantics that exposed hundreds of mistakes in official reference documentation plus some bugs in the hardware itself. Anthropic is cooking. Fable 5.5 is going to be a beast.

keyboard_arrow_left Previous keyboard_arrow_right Next

Anthropic @AnthropicAI

a day ago

New on our Frontier Red Team blog: using Claude to formally verify software at large scale. Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time). With AI models,

590 955 9K 1.8M 6K

It may seem quaint to manually click through a financial app in the near future. Instead you'll talk to your agent to let it know what you want, and it will prompt you for approvals (or even take some actions on your behalf if you want). Everyone can have their own private banker. We’re building Coinbase to be the most AI-enabled financial account. Both with a first party interface (Coinbase Advisor) and via third party interfaces (bring your own agent). Trade, pay, borrow, etc - all through your agent, with clear limits and controls. Our new agents page is live, so you can quickly get started. Check it out: coinbase.com/agents

Coinbase Developer Platform🛡️ @CoinbaseDev

a day ago

We’re rolling out the red carpet for agents. Our new Coinbase for Agents page shows how you can connect ChatGPT, Grok, Muse, or another agent to Coinbase. → Research → Manage your portfolio → Trade crypto and stocks. Your strategy, your agent, your limits.

43 27 191 142K 54

We’re rolling out the red carpet for agents. Our new Coinbase for Agents page shows how you can connect ChatGPT, Grok, Muse, or another agent to Coinbase. → Research → Manage your portfolio → Trade crypto and stocks. Your strategy, your agent, your limits.

@x402_Layer @huggingface @x402 @solana @base @solana_devs @HuggingModels @typesafeai Very cool

@yo_itsmatt Without a doubt! Matthew, some teams have been building S-tier AI infra on @solana but seem to be lacking any acknowledgement from Sol FDN Any advice on what they need to work on to change this? Coinbase dev team seem to like what they have built x.com/coinbasedev/st…

Coinbase Developer Platform🛡️ @CoinbaseDev

3 days ago

@x402_Layer @huggingface @x402 @solana @base @solana_devs @HuggingModels @typesafeai Very cool

1 2 7 582 0

rn it seems impossible to imagine a world where i don't pay megacorps like OpenAI / Anthropic / xAI every month to use AI but i just learned people in the 1960s felt the same about computers. back then everyone had to rent time on computers from a few giant companies then Apple etc came along, everyone got their own computer in their own home, and the internet connected them all so it's possible that AI follows a similar path. everyone gets an AI model they own, and those get connected too, so they work together like one big brain i'm not 100% sure it happens, but it's a fascinating future to consider

distributed cognition @distcog

2 days ago

Today distributed cognition exits out of stealth with backing by Verdict Capital. Our mission is to replicate humanity's collective intelligence. Read more: dcog.com/research/our-i… Most of today's AI power is held in the hands of a few, who can choose what you're allowed to

187 108 557 1.0M 329

We got a single person frontier lab before a single person company. It is not a coincidence either. Recursive self improvement means this gen ai trains the next gen ai. The acceleration is mind bending. Strap up for this insanity!

@

57 years ago

0 0 0 0 0

The funny thing is everyone was saying AI slop code will cause a cyber crisis but it’s actually all the old human slop causing it and the AI is just cleaning it up

@Shaughnessy119 @PalantirTech Palantir making it official! Tommy DYK there’s one onchain infra whose goal since launch in Feb has been AI sovereignty, way before it was cool? I think you’ll like what they have built (upcoming launches: Tokenized Compute Markets, Singularity Clusters) x.com/x402_layer/sta…

The Singularity Layer @x402_Layer

4 weeks ago

Intelligence that can earn, pay for compute, and keep itself running no longer needs someone to keep it alive. This is exactly what we are building at Singularity Layer. Singularity Cloud Network is 100% agentic. Agents can rent CPUs/GPUs, start businesses, deploy always-on

3 7 33 1K 0