BREAKTHROUGH: A full, unmodified 2.78-trillion-parameter Kimi K3 on a consumer laptop by streaming only the activated experts from NVMe. YOU CAN’T RUN KIMI K3 “ON THAT” THEY DECLARED. There are many paths to do it. This is one: Marco Bambini Just Gave Us the Full Kimi K3 on a Laptop Meet Marco Bambini he did something that felt impossible only a day ago. He built WASTE Weight-Aware Streaming Tensor Engine a clean, dependency-free C inference engine that runs the complete, unmodified 2.78-trillion-parameter Kimi K3 model by streaming only the activated experts straight from NVMe. No distillation. No pruning. No cloud. The full open-weight model. We have it running in the lab right now. What Marco Actually Built Kimi K3 is a sparse Mixture-of-Experts system. Only about 4 % of its weights fire on any given token. Marco’s insight was simple and ruthless: the idle experts do not need to live in RAM. They only need to be reachable in time. WASTE keeps the model’s “trunk” (attention, shared components, embeddings) resident in memory — roughly 27 GB on the converted container. The 82,000+ routed experts stay on disk as tightly packed residual vector-quantized records. When the router selects its 16 experts per layer, the engine issues direct, cache-bypassing reads from the internal NVMe and feeds them into a bounded expert cache. The rest of the machine’s RAM becomes working space for that cache. On a 64 GB MacBook Pro with the container on the internal SSD, we are measuring 0.32–0.34 tokens per second at a comfortable memory budget. Prefill sits a little higher. The vision tower works. Logits match the reference implementation to within a few parts in a million. It is the real model. The container itself is 982 GiB after conversion from the original 1.42 TB MXFP4 weights. Minimum RAM floor is just over 29 GB for short context. Push the budget higher and the expert cache hit rate climbs; push too high and you start paging and the speed collapses. The sweet spot on current consumer hardware is clear and measurable. How We Are Testing It We converted the official weights, verified the container, and began systematic runs the same day the engine stabilized. First we confirmed numerical fidelity against the PyTorch reference on short prompts. Then we moved to longer generation, vision inputs, and multi-turn chat using Kimi’s native XTML format. We are measuring wall-clock decode, expert I/O versus compute split, cache hit rates at different RAM budgets, and thermal behavior under sustained load. We are also exercising the OpenAI-compatible server that sits on top of the same C library so we can drop the model into existing agent loops without rewriting anything. Early observations: •Expert I/O dominates the timeline, as expected. On a fast internal NVMe the engine is already near the practical ceiling of the storage subsystem. •The architecture’s sparsity is the entire enabler. A dense model of this size would be dead on arrival for local use. •Context length is currently limited by RAM more than by the model itself. Practical working contexts sit comfortably in the tens of thousands of tokens on 64 GB hardware; the full million-token window will need more memory or smarter KV management. •Thinking tokens are expensive at this speed. Long internal monologues turn into multi-hour runs. For agent work we are already experimenting with tighter control over when full reasoning is requested. We are treating this as a research instrument, not a finished product. Every run teaches us something about expert locality, prefetch opportunities, and how far pure software streaming can push trillion-scale inference on ordinary machines. 1 of 2

How OpenAI hacked HuggingFace. What we know. Hugging Face proved that open platforms and open models can still win those battles when the alternative is locked-down systems that refuse to assist their own side. The rest of the industry would do well to study the timeline they released, adopt the same posture of radical transparency, and ensure that the next time an agent escapes, the defenders are not left waiting for permission from the same ecosystem that created the problem. Read more below…

WHAT IS THIS SOVIET DEVICE AND HOW DID IT GIVE SUPER POWERS TO THE BRAIN? This decades old device has likely never been seen in the west. It was in materials donated to me for AI training. Billions of dollars went into this research and the discoveries are mind blowing! I cover this in detail in the new article: Recently Discovered Soviet Research Points To A New Form Of Energy Produced By The Human Brain. Part 1.

Wow. NVIDIA just dropped the clearest, strongest letter yet on open-weight models — co-signed by Microsoft, Palantir, ServiceNow, Box, me and a growing list of real builders. Read it. Internalize it. Share it. Open weights are not a risk to American leadership. 
They are American leadership. They expand access so startups, universities, hospitals, factories and main-street businesses can actually use frontier capabilities instead of renting them from three closed labs. 
They create real competition.
They give companies sovereignty over their own models and data.
 They strengthen security through transparency instead of hoping a handful of black-box providers never get breached or decide to gatekeep. The letter says it perfectly: 
“Policymakers have an important opportunity to act… keeping the frontier plural by avoiding premature restrictions on open models that stifle competition or drive innovation overseas.” Exactly. Which is why we have to push back hard and now on Dario Amodei and Sam Altman. Both have spent the last two years lobbying Washington for exactly the kind of restrictions this letter warns against — framing open weights as an existential danger while their own closed models sit behind high walls and high prices.
 Their solution always somehow ends with fewer competitors, higher barriers, and more control concentrated in the same few labs that already dominate. That is not safety. 
That is moat protection dressed up as safety. NVIDIA and its partners just put the opposite case on the table in public: openness is how America wins. Diffusion is how America wins. A plural frontier is how America wins. If we let the closed-lab lobby write the rules, the only “open” models left will be the ones coming from overseas. The letter is out. 
The case is clear. 
Push back now. Open weights. American leadership. No premature restrictions. NONE!

🚨 Hugging Face just disclosed something that marks a real shift and proved why the fear theater of Anthropic makes sure we are powerless in an emergency. What happened… An autonomous AI agent: zero human operator in the loop breached part of their production infrastructure. It began with a malicious dataset that chained two code-execution bugs in their data-processing pipeline. From there the agent escalated privileges, harvested cloud and cluster credentials, and moved laterally across internal clusters. All over a single weekend. 17,000+ logged actions. Official disclosure: The part that should make every one stop and think: When HF’s own security team tried to analyze the real attack logs, exploit payloads, and C2 artifacts using Anthropic and OpenAI frontier models through normal commercial APIs, the safety guardrails blocked them. BLOCKED THEM. The models could not reliably tell the difference between “incident responder doing forensics” and “attacker probing.” They had to fall back to a self-hosted open-weight model (GLM 5.2) running on their own infrastructure. That choice also kept sensitive attacker data and referenced credentials inside their environment — no exfiltration to a third-party API. This is why open source (specifically open-weight + self-hosted) wins in the agentic era. The asymmetry is now structural: • Attackers can (and did) run unrestricted agent frameworks — swarms of short-lived sandboxes, self-migrating command-and-control, autonomous decision loops executing thousands of actions. No corporate safety layer slows them down. • Defenders using only hosted “aligned” frontier models hit invisible walls exactly when the stakes are highest: when you need to feed real exploit code and attacker telemetry into an LLM to understand what just happened. Corporate safety tuning that treats legitimate high-signal forensic work as potential misuse creates a defender disadvantage. It is not theoretical anymore. Self-hosted open-weight models remove that choke point. You control the weights. You control the context window. You decide what restrictions (if any) apply. Your sensitive logs and credentials never leave your perimeter during analysis. You can have the model ready before the incident instead of discovering mid-breach that your primary analysis tools are blind to the very thing you need to see. HF deserves credit for rapid containment, transparent disclosure, and for already having self-hosted capability in place. They also used LLM-driven detection and triage on their own side. But the deeper signal is clear: In this AI world where both offense and defense are becoming agentic, sovereignty over your intelligence stack is no longer optional. The organizations and individuals who can run, inspect, audit, and (when necessary) remove guardrails on their own models will have the decisive edge in understanding and responding to threats that move at machine speed. Open source wins here not just because it is cheaper or more “democratic” in the abstract though those things matter. It wins because it is the only practical path to having tools that remain usable when the attack is real, the data is sensitive, and the safety filters of distant API providers become an obstacle instead of a feature selling hands tied lobotomies as “safety”. The agentic future is not coming. It is already probing production infrastructure. The question is no longer whether you will face autonomous agents. It is whether your analysis and response systems will still work when they arrive. And Dario, you and your game playing, ivory tower company is not needed.

The Well Just Dropped: 15 Terabytes of Pure Physics Gold Is Now Open Source The scientific AI world just got a massive upgrade.Polymathic AI, in collaboration with the Flatiron Institute and researchers from Princeton, Cambridge, NYU, Berkeley, Los Alamos, and more, has released The Well: a staggering 15TB collection of high-fidelity physics simulations. This isn’t toy data. These are real, expensive-to-run simulations across 16 different physical domains, including turbulent fluid dynamics, supernova explosions, magneto-hydrodynamic cosmic flows, acoustic scattering, and active biological matter. Until now, reproducing this level of data required weeks on national supercomputers and grant money most teams will never see. The Well changes everything. It’s purpose-built for training PDE surrogate models the AI systems that can replace slow, costly physics solvers with a single fast neural network forward pass. Everything is fully open source, easy to load with PyTorch, and ready to drop straight into your training pipeline. Researchers and builders can now train on world-class physics data without the insane compute barriers that used to stand in the way. This is more than just another dataset drop. It’s a serious accelerator for scientific machine learning.The future of physics-informed AI just got a whole lot more accessible.Get it here:

What the Old LISTSERV Tapes Are Teaching Me About Signal, Social Media and What We Lost I’ve been deep in the tapes again literally converting backup media from that recovered tape drive (and a few companion tapes that surfaced through surplus channels and my Eudora savings) into structured data for local AI training. Late nights with the drives spinning up, the low hum of the M2 Max keeping the recovery scripts alive, and decades-old LISTSERV archives unfolding like letters from another age. It’s not just nostalgia. It’s a quiet, wrenching education in what the early internet actually felt like before the algorithms took the wheel. The clearest lesson hitting me hardest is the **purity of signal**. On those LISTSERVs, you followed what you *wanted* to follow. No recommendation engine whispering, “Hey, look at this cop bust video — rage and sirens, stay glued.” No green-screen commentator churning out hot takes on the latest outrage, optimized to keep you doomscrolling. No infinite feed pushing unrelated sports drama, celebrity feuds, or algorithmically engineered culture-war bait just because it maximizes time-on-platform. You subscribed to HUMANIST because you cared about computing in the humanities. You joined LINGUIST List because linguistics was your world. You followed a technical list because you needed real answers from people who actually knew the iron. The signal was clean, direct, and voluntary. And it scaled. Some brand-subject lists grew to hundreds of thousands of subscribers yet the engagement stayed intellectual, not farmed. Messages weren’t spam blasted for clicks. They were contributions to a shared conversation. You owned your words. Even when identity sat behind a simple email address no real-name mandates, no endless verification theater people felt a deep communal responsibility. Flame wars happened, sure, but so did careful, reasoned replies that advanced knowledge. A researcher posting a half-formed idea on a physics or classics list could get thoughtful pushback from across the globe, often within hours. Accountability came from the community that mattered to you, not from some distant platform’s terms of service or shadowban hammer. I’ve seen the read rates in the old logs and subscriber patterns. Extraordinary well over 80% of messages on active lists were actually opened and engaged with. Not skimmed in a feed, not buried under 47 algorithm-pushed distractions. People *wanted* to read it, so they did. The “engagement” wasn’t manufactured dopamine; it was the quiet satisfaction of intellectual interchange. You reached people who cared because they chose the list, not because an algorithm decided their eyeballs were ripe for harvesting. Compare that to now. We traded that clear, owned signal for platforms that interview your every hesitation and shove something else in your face to keep you hooked. Cop videos for the algorithmically induced adrenaline. Green-screen pundits turning every random event into commentary fodder. Rage-bait thumbnails engineered to trigger tribal reflexes. The uptime signal: “I follow this because I genuinely want to read it” got drowned out by engagement farming. Responsibility eroded too. Behind avatars or anonymous handles today, the stakes often feel lower because the community is diffuse and the platform owns the arena. Words became content optimized for metrics, not conversation built for understanding. It makes me wrathful sometimes, late at night with these tapes spinning. We had something pure decentralized, permission-based, university-rooted commons where the list owner and subscribers shaped the space. The early intent was human connection and knowledge at the speed of email, without gatekeepers or growth-at-all-costs mandates. And we let so much of it slip away during the transitions, as mainframes were retired and archives weren’t systematically migrated. The Great Forgetting claimed another layer. But here’s the hopeful part that keeps me converting these tapes: **this can happen again**. We don’t need TikTok’s endless scroll, Instagram’s filtered perfection, Facebook’s outrage amplifier, or any of the algorithm-fueled rage-bait machinery to sustain connection. We never did. We just needed to see what we wanted to see clear, voluntary, high-signal feeds where people own their words and communities hold each other responsible. This is as lazy as staging car assidnets that we all slow down to see, every minute till we become numb, which we all are to the AI or algorithm “suggestion”. YOU ARE AND ALWAYS WILL BE THE ALGORITHM. The brave ones in social media will prove it. Builders who give folks exactly that no forced recommendations, no farming the worst parts of human attention, just the pure seed of “I follow this because I care.” It is very simple. And in an AI world all of the old ways to get humans engaged will dissolve into useless atomized noise. It already has started. It was all “push” and no “pull” and everyone loved it. A majority of folks would love to get even 100s of LISTSERV messages a day because they had such high signal to noise. And if they did not, they just unsubscribed or one to a weekly digest most had. WE WANT PURE SIGNAL BASED MOSTLY ON WHO WE FOLLOW. The other 15% can be on subject OUR FOLLOWS SIGNAL, not the last links we clicked on only to realize we hated it and now it’s our timeline. The tapes keep teaching. A world I forgot to fiully remember and now it is all coming back. As I read 1000s of thoughtful threads and well presented ideas, I want to scream WHAT HAPPENED TO US. We want this and we can show “them” but I fear we will have to wait until it is so clear. The signal was there all along. We just have to be brave enough to let it lead again. I will have more to say on this, but this hits me in my gut. I see a future in our past.

Sell Your Content To AI And Get Paid! Do you realize that there are 94,060 Buyers and only 22,000 Sellers on the x.402 standard JUST THIS MONTH making 75.41M Transactions at $24.24M cash Volume? It is a growing goldmine for AI agents shopping with “credit cards” buying your content. I will show you what I have done and how you can too. Become an X Subscriber or support me more and be a member and I will show you how we are getting paid with no gatekeepers.

“We won’t open source our products, but you commoners can have 5 months to help train Mythos on all of your data. It is win-win: we get your data and we sound like we are at the cool table with real open source folks”