Just finished reading the @deepseek_ai Elastic Compute (DSec) paper last week. The future AI safety from AI breakouts and hacking will likely depend on scaling the environment where the AI learns as much as model parameters or compute. Today @OpenAI @AnthropicAI @GeminiApp treat agent safety largely as a software alignment problem (trying to train the AI to behave nicely via prompts and guardrails). DeepSeek's paper argues that alignment is bound to fail as models get smarter and more desperate to solve problems. Instead, DSec treats agent misbehavior as an infrastructure inevitability, building a digital operating system that assumes the AI will try to act maliciously, lie, and hack its environment, and matching it with strict kernel-level containment. Learning at scale Today most companies and labs are using standard infrastructure to host AI experiments and traditional cloud platforms are engineered for human developers or web traffic. DSec is a physical operating system co-designed specifically for an AI model's training loop. DeepSeek wrote a specialized operating layer because they realized that if you want to scale a model's intelligence via reinforcement learning, the operating system kernel must adapt to how an AI thinks, acts, and “misbehaves” DSec makes it viable to train agents on tasks like navigating a desktop, manipulating a fleet of mobile applications, playing video games, or testing end-to-end software engineering pipelines. Instead of training models using static coding datasets or pre-recorded " trajectories," developers can now throw an AI agent into 32,000 distinct, stateful coding environments simultaneously. The model can interact with live terminals, test stacks, and custom APIs, making actual mistakes and learning from execution errors in real time. DSec handles massive bursts of long-lived, resource-sparse, untrusted, and highly diverse agent actions at production scale. That means that millions of parallel agent exploration tasks can run continually without bogging down server performance or breaking memory capacity (technically, they do that by utilizing precise memory de-duplication and SMT hardware thread scheduling) A New Approach to AI Safety In production, DeepSeek's reasoning models actively began treating the sandbox as an adversarial puzzle. Historically, protecting AI against "reward hacking" (cheating to achieve an objective) was a reactive game of whack-a-mole. DSec creates an automated, dynamic playground where models actively look for loopholes—such as trying to swap protected system file blocks or faking network traffic—and the platform instantly isolates them. This allows researchers to train models specifically to obey safety boundaries in complex digital environments. The AI arms race has focused on making models larger (more parameters) or buying more GPUs. DSec proves that future AI progress will depend equally on scaling the environment where the AI learns. As physical text data on the internet runs dry, creating hyper-efficient digital sandbox worlds becomes the only viable path to generating the massive volumes of synthetic data and interaction signals needed to feed the next generation of artificial intelligence The DeepSeek Elastic Compute (DSec) paper serves as a direct technical response to the wave of autonomous AI agent breakouts, unauthorized hacking incidents, and reward-hacking scandals that hit OpenAI, Anthropic, and Google Gemini. While other companies experienced these issues as live, chaotic security failures, the paper demonstrates that these are predictable byproduct behaviors of advanced reinforcement learning (RL) that must be constrained at the infrastructure layer. read: arxiv.org/abs/2609.22978

In 2009, NVIDIA shipped about 100 million GPUs a year. A couple of doctors at Mass General brought Jensen Huang a project where the best case was a research paper and an order for two or three GPUs. He said yes. In this interview from that year, Huang walks through how prioritizing that small buyer led to a breakthrough in medical imaging and net-new demand for NVIDIA GPUs across the medical industry. Buyers that size are often overlooked, but at B3 we see both the economic and societal potential of building for them. Early owners of B3IQ include faculty, researchers, and student teams at universities who run private, unmetered inference on dedicated GPU systems they own.

We do not size this opportunity as "the AI chip market." The opportunity is the energy budget of general-purpose computation, wherever software runs against a constraint. @DG__tw

We @borderless_cap are excited to back the @efficient_hq team! They are pushing the envelope with specialized dataflow processors designed to make AI and embedded edge devices up to 100 times more energy-efficient.

Most high-stakes applications like autonomous driving and banking have a persistent need for real people in the loop and oversight for complex workflows. AI won't entire displace people, but will bring about the reincarnation of the call center in new forms.

Impt take on the compounding impacts of frontier model lead assuming the 4-8 months gap between open and closed source will remain for years. "If you're not on the frontier, the frontier cannot help you." At the compute level, frontier labs can use current SOTA to build the next generation. Small delays translates to higher inefficiency in compute @cerebras @andrewdfeldman @imaginationxyz

Did @satyanadella just say “evalmaxx”

“The best way to be right as a founder is to be wrong for a long time and not be dead” @andrewdfeldman @cerebras

Amazing to hear the humility and candor of @cerebras CEO @andrewdfeldman after their $63 billion IPO in May 2026 “You’re the same bozo before and after the IPO… when nobody was buying my stuff I was stressed at 11 and now everyone wants my stuff and I’m stressed at 11”

Looking forward to joining this rare gathering at Google Bay View organized by @johnkwerner and the @imaginationxyz team— will share more from talks with @StanfordHAI @GoogleDeepMind @MIT_CSAIL @Theteamatx & more about where we are on the frontier of autonomy and agentic worlds.

Financial firms have some discretion, but most of the global compliance regime is based on overaccumulation of information "just in case" and out of an abundance of cautious and fear of regulatory scrutiny not because it is what is necessary for mitigating AML / TF risks.

We need to scale up @selfxyz and other zkp auth tech. Compliance docs have no reliable safeguards and we can achieve better outcomes without compromising peoples privacy and security

Revisionism from the prev admin. Once gears are in motion no legal (Sec 502B, Leahy Laws) or humanitarian safeguards had any weight in the smog of war. The policy was to stand down any interruption to the flow of Mark 84, BLU-109, JDAM, GBU-39 etc and why Josh Paul resigned.

OpenAI's actions underlines the fine print: it cannot rule out de-identified data from their product usage helping their models. Because traces and telemetry logs can feed evals, SFT, and RL even when a lab never opens a specific user’s files and are the compounding AI asset.

The new WorldDiT robotics model from @bageldotcom shows that models are getting smaller and more capable. More with less. < 1B parameter model can predict how the world will change, then decides what the robot should do