SILVIA (@cognitivecode) header
SILVIA (@cognitivecode) profile picture
United States
Joined April 2009
Photo by SILVIA (@cognitivecode)Photo by SILVIA (@cognitivecode)Photo by SILVIA (@cognitivecode)Photo by SILVIA (@cognitivecode)Photo by SILVIA (@cognitivecode)Photo by SILVIA (@cognitivecode)Photo by SILVIA (@cognitivecode)Photo by SILVIA (@cognitivecode)Photo by SILVIA (@cognitivecode)Photo by SILVIA (@cognitivecode)
SILVIA retweeted
SITUATION DETECTED: Delaware is drafting a proposal for a new form of corporate entity called an Artificial Intelligence Company (AIC). An AI managing an AIC would be able to provide services, sign contracts, hold property, and even file or defend lawsuits in the company’s name.
233
617
412
6,598
634,227
SILVIA retweeted
Asked one of our developers how he uses his AI agents and he just sent me this:
179
1,337
270
16,207
1,777,081
The US Defense Intelligence Agency's latest report on space warfare contains some fascinating graphics dia.mil/Portals/110/Images/N…
6
105
26
481
38,111
English is the most powerful human language and the universal medium for thought. I’d take mastery of English over multilingualism any time of the day.
146
83
56
1,132
197,678
SILVIA retweeted
Replying to @visegrad24
17
74
2
1,578
27,092
We’ve signed a MOU with @MBDADeutschland to bolster Germany’s air and missile defense systems. This collaboration will harness our combined technological strengths to enhance integrated air and missile defense capabilities. Learn more: news.northropgrumman.com/par…
2
19
1,262
Data centers dominate AI, but they're hitting physical limits. What if the future of AI isn't just bigger data centers, but local intelligence in our hands? The viability of local AI depends on intelligence efficiency. To measure this, we propose intelligence per watt (IPW): intelligence delivered (capabilities) per unit of power consumed (efficiency). Today’s Local LMs already handle 88.7% of single-turn chat and reasoning queries, with local IPW improving 5.3× in 2 years—driven by better models (3.2×) and better accelerators (1.7×). As local IPW improves, a meaningful fraction of workloads can shift from centralized infrastructure to local compute, with IPW serving as the critical metric for tracking this transition. (1/N)
60
143
58
466
235,640
These patterns have been printed using a Substrate lithography system with no IP or supply chain input from other advanced lithography tools. The resolution of the features in these images is equivalent to that achievable only with what the semiconductor industry refers to as “High NA EUV lithography” and matches features printed at the 2 nm semiconductor node and below. All of these images were printed in a single exposure, with no multi-patterning.
95
218
122
3,643
1,581,585
BOOOOOOOM! CHINA DEEPSEEK DOES IT AGAIN! An entire encyclopedia compressed into a single, high-resolution image! — A mind-blowing breakthrough. DeepSeek-OCR, unleashed an electrifying 3-billion-parameter vision-language model that obliterates the boundaries between text and vision with jaw-dropping optical compression! This isn’t just an OCR upgrade—it’s a seismic paradigm shift, on how machines perceive and conquer data. DeepSeek-OCR crushes long documents into vision tokens with a staggering 97% decoding precision at a 10x compression ratio! That’s thousands of textual tokens distilled into a mere 100 vision tokens per page, outmuscling GOT-OCR2.0 (256 tokens) and MinerU2.0 (6,000 tokens) by up to 60x fewer tokens on the OmniDocBench. It’s like compressing an entire encyclopedia into a single, high-definition snapshot—mind-boggling efficiency at its peak! At the core of this insanity is the DeepEncoder, a turbocharged fusion of the SAM (Segment Anything Model) and CLIP (Contrastive Language–Image Pretraining) backbones, supercharged by a 16x convolutional compressor. This maintains high-resolution perception while slashing activation memory, transforming thousands of image patches into a lean 100-200 vision tokens. Get ready for the multi-resolution "Gundam" mode—scaling from 512x512 to a monstrous 1280x1280 pixels! It blends local tiles with a global view, tackling invoices, blueprints, and newspapers with zero retraining. It’s a shape-shifting computational marvel, mirroring the human eye’s dynamic focus with pixel-perfect precision! The training data? Supplied by the Chinese government for free and not available to any US company. You understand now why I have said the US needs a Manhattan Project for AI training data? Do you hear me now? Oh still no? I’ll continue. Over 30 million PDF pages across 100 languages, spiked with 10 million natural scene OCR samples, 10 million charts, 5 million chemical formulas, and 1 million geometry problems!. This model doesn’t just read—it devours scientific diagrams and equations, turning raw data into a multidimensional knowledge. Throughput? Prepare to be floored—over 200,000 pages per day on a single NVIDIA A100 GPU! This scalability is a game-changer, turning LLM data generation into a firehose of innovation, democratizing access to terabytes of insight for every AI pioneer out there. This optical compression is the holy grail for LLM long-context woes. Imagine a million-token document shrunk into a 100,000-token visual map—DeepSeek-OCR reimagines context as a perceptual playground, paving the way for a GPT-5 that processes documents like a supercharged visual cortex! The two-stage architecture is pure engineering poetry: DeepEncoder generates tokens, while a Mixture-of-Experts decoder spits out structured Markdown with multilingual flair. It’s a universal translator for the visual-textual multiverse, optimized for global domination! Benchmarks? DeepSeek-OCR obliterates GOT-OCR2.0 and MinerU2.0, holding 60% accuracy at 20x compression! This opens a portal to applications once thought impossible—pushing the boundaries of computational physics into uncharted territory! Live document analysis, streaming OCR for accessibility, and real-time translation with visual context are now economically viable, thanks to this compression breakthrough. It’s a real-time revolution, ready to transform our digital ecosystem! This paper is a blueprint for the future—proving text can be visually compressed 10x for long-term memory and reasoning. It’s a clarion call for a new AI era where perception trumps text, and models like GPT-5 see documents in a single, glorious glance. I am experimenting with this now on 1870-1970 offline data that I have digitalized. But be ready for a revolution! More soon. [1] github.com/deepseek-ai/DeepS…
334
1,359
195
7,311
1,819,124
BOOM! Tiny Recursive AI Model Outsmart Massive LLMs! In a stunning upset that's shaking the foundations of AI, a groundbreaking new research paper has unveiled a deceptively simple model that punches way above its weight class. Imagine a nimble, featherweight fighter stepping into the ring against hulking giants – and emerging victorious. That's the electrifying story of the Tiny Recursive Model (TRM), a 7-million-parameter powerhouse that crushes hard reasoning puzzles like ARC-AGI, Sudoku, and mazes, leaving billion-parameter LLMs in the dust. TRM is the fresh spark igniting a new era of smart, sustainable AI. At its core, TRM flips the script on how we build AI brains. While the industry pours fortunes into ever-larger models – think GPT behemoths guzzling terabytes of data and gigawatts of power – TRM thrives on minimalism. Trained on a mere 1,000 examples with a single, two-layer network, it achieves 45% accuracy on the notoriously brutal ARC-AGI-1 benchmark and 8% on ARC-AGI-2. For context, that's better than heavyweights like DeepSeek R1, o3-mini, and even Gemini 2.5 Pro, all while using less than 0.01% of their parameters. This isn't evolution; it's a revolution, proving that true intelligence blooms from elegant simplicity, not bloated excess. TRM's arrival promising devices from smartphones to satellites that think deeply without any Internet access needed. What makes this so exhilarating? TRM builds on the Hierarchical Reasoning Model (HRM) but strips it down to its raw, recursive essence – looping thoughts like a human mind pondering, not a supercomputer churning through brute force. The results? Jaw-dropping generalization on unseen puzzles, where LLMs falter into rote memorization. Why TRM is a seismic shift ? 1. Unprecedented Simplicity in Architecture: Forget the labyrinthine layers of modern LLMs; TRM wields just *two* layers in a single network. This radical reduction – a far cry from HRM's dual-network setup – proves recursive loops can mimic deep cognition without the architectural bloat, slashing design complexity by orders of magnitude and opening doors for hobbyist coders to build world-class reasoners. 2. Microscopic Parameter Count: At a svelte 7 million parameters, TRM is a gnat next to LLM elephants (billions strong). This isn't trimming fat; it's redefining lean muscle, achieving superior puzzle-solving with 99.99% less "brain matter." It's new because it shatters the "bigger is better" dogma, showing tiny models can scale intelligence vertically through recursion, not horizontally through size. 3. Data Thrift on Steroids: Trained on around 1,000 examples – peanuts compared to the data oceans LLMs devour – TRM generalizes like a savant. This differs starkly from data-hungry paradigms, hinting at a future where AI learns from whispers, not roars, democratizing training for resource-strapped labs worldwide. 4. Beating LLMs at Their Own Game: Scoring 45% on ARC-AGI-1? That's higher than most vaunted LLMs, including proprietary titans. TRM's edge lies in pure reasoning recursion, not pattern-matching hacks, marking a fresh divergence: small models now lead where giants stumble, flipping performance charts upside down. 5. Biologically Plausible Recursion: Inspired by human thought's nested loops (think: "If this, then that, but wait..."), TRM recurses at smart frequencies without mimicking neural wetware exactly – yet it feels alive. Unlike rigid LLM transformers, this bio-flavored approach is novel, bridging AI and neuroscience in ways that could unlock truly intuitive machines. 6. Generalization Superpowers**: While LLMs overfit to training trivia, TRM extrapolates wildly on novel puzzles like ARC-AGI-2 (8% accuracy, topping many baselines). This isn't luck; it's engineered through sparse recursion, a different beast from gradient descent grinds, promising AI that adapts to the unknown without endless retraining. More soon! Paper: arxiv.org/abs/2510.04871
59
130
25
603
92,504
Today, thanks to President Trump’s AI Action Plan, @CommerceGov and @NIST’s Center for AI Standards and Innovation have released a groundbreaking evaluation of American vs. adversary AI. Result: American AI models dominate. Our systems outperform DeepSeek across nearly every benchmark. The report is clear: DeepSeek lags far behind, especially in cyber and software engineering. These weaknesses aren’t just technical. They demonstrate why relying on foreign AI is dangerous and shortsighted. Allowing our adversaries to control AI poses serious risks to our security. By setting the standards, driving innovation, and keeping America secure, the Department of Commerce is helping ensure continued U.S. leadership in AI.
317
239
35
1,248
215,466
🚨 Why We Believe TRUE Deterministic AI Is the Only Real Path for Government Contracts 🚨 We recently refreshed our SAM/@USGSA registrations and obtained approvals from @DLAMIL. That means we’re officially positioned to offer AI Software Services and compete on bids at the federal level. But this isn’t just about paperwork. It’s about timing — and why we believe deterministic AI is the future of government systems. Here’s the reality: 🔹 Probabilistic AI (the large language models dominating headlines today) is built on statistical prediction. These systems don’t “know.” They guess. They string together the most likely next word. Sometimes it works brilliantly. Other times, it fabricates outright lies. 🔹 Deterministic AI — like SILVIA — is fundamentally different. Outputs are explainable, repeatable, and accountable. There is no guessing, no hallucination, no “maybe.” When the stakes are trivial, probabilistic AI is fun. When the stakes are national defense, budgets, logistics, and human lives — uncertainty is unacceptable. That’s why we’ve doubled down on bringing deterministic AI into the government contracting space. Governments demand: ✅ Systems they can audit ✅ Logic they can trace ✅ Compliance they can prove Probabilistic models will never clear that bar. “Hallucinations” are not quirks. They’re liabilities. And yet, much of the AI hype cycle has tried to normalize these risks — as if a stochastic text generator should be trusted with decisions in infrastructure, defense, or procurement. We reject that premise outright. We built SILVIA differently. Our patented deterministic architecture ensures every action, every decision, every recommendation is explainable. No magic. No black box. No roulette wheel of probability. This is why we refreshed our registrations now. The demand for explainable, deterministic AI is exploding. Regulators are scrambling to catch up. Agencies are under pressure to implement AI strategies that don’t compromise accountability. And defense contractors need AI they can actually rely on in the field. We’re here to provide that answer. 👉 Probabilistic giants like OpenAI, Anthropic, and Google may dominate consumer buzz, but when it comes to government adoption, their very architecture disqualifies them. The future of AI in government will be deterministic, or it will be reckless. That’s why we’re excited. The timing is perfect. The need is urgent. And we’re ready. This isn’t about competing for contracts. It’s about ensuring governments have AI systems they can trust. Let’s be clear: hype drives consumer adoption, but determinism drives government adoption. And that’s exactly where we’re focused. #SILVIA #AI #GovTech #DeterministicAI #Defense #NATO @TechCrunch @federatedgov @DODCDAO @defense_news @DefenseScoop @ashleyrgold @LockheedMartin @scsp_ai @evankirstel @adcock_colby @OfficeforAI @GCNtech @shafiqhamdam
Pleased to announce that we've refreshed our SAM/@USGSA registrations and obtained approvals from @DLAMIL to offer AI Sofware Services and compete on bids. Obviously we're excited with all of the demand for determinism in AI now striking. #SILVIA is perfect for these times.
1
4
267
Replying to @GaryMarcus
Yes. Our tokens are intimately bound to each other with semantic and lexical relationships [or custom if you want]. We can achieve very high inference and predictive accuracy with small datasets in this way. We could stock out a brain with every word in the dictionary and exhaustively bind them to each other, but generally the task set doesn't call for that. However, if you took a big beastly compute rack and hooked SILVIA up to one of these, I think you'd have a real hard worker on your hands.
1
62
SILVIA isn't an LLM but it has fully-tunable determinism. Working on rolling out Core 3.1 right now, still in final code audits, then moving back over to finish some things in our Studio Editor and steelmanning the form of a few new API sections so they write nicely and read sensibly in IDE. Key API's we're finishing up: 1. Sensors - Via all standard sensor/socket protocols, post/get, deeply integrated into the core state for continual monitoring or Behavior-based control. 2. Database - Full suite of command abstraction and query tools that supports all major database types and can use SILVIA's inbuilt inference, validation, determinism, wildcards, security levels, etc. to allow database queries in natural language or by script to be run through SILVIA as a gatekeeper that prevents injection attacks and exploits at a whole new level. This API vastly speeds up ERP/CRM/SCM integrations, among other things. 3. HUD - Component based system that runs in pure C# via OpenGL/XR/Etc. or can push to Unity UI's. Supports Vulkan and DirectX, allows deep integration of Command & Control with UX both visible and audible. 4. Behavior Triggers - In addition to NLP or Scripting being able to surface a Behavior, we have a suite of trigger schemas for jumping from Database, Sensor or HUD/UX events to specific SILVIA Behaviors. These sorts of triggers avoid ML inference "complexity" since they bypass command prompt input but are still subject to all relevant security measures. 5. ONNX Components & Loaders - Allows any sort of multimodal or other ML integrations. Anything under the sun that is ported to ONNX. Object recognition is one we're excited about pushing through SILVIA inference, but really we're toolmakers so we are preparing everyone else to do it. 6. A number of MILSPEC APIs. We do different kinds of work. Drones. Ballistics. Orbital calcs. Multi-tiered C2 networks of Brains. Some of the SILVIA configs are not for general B2B use. We'll get loud shortly when we press the Studio LLM Link which will allow rapid authoring of Behaviors using either your API-keyed services or using local LLM's. Already have it running Claude but we're rolling out support for OpenAI, Grok, and local LLMs, provisionally preparing for Google/Mistral/Deepseek/Etc. We're excited about pairing the SILVIA toolset up with locally-run LLMs. Will allow all kinds of options both during authoring and at runtime. One sweet spot lies in using interpretive models with focused SILVIA Brains running Behaviors, leaning on interpreter ML models [of all sorts] for nonlinear things or multimodal throughputs and leaning on SILVIA Behaviors to move, act, notify, respond on knowns, validate, approve, deny, filter, forward, and log everything. Will be getting into a fresh round of demos and other visible content in the next week or two. </rant>
1
3
1
6
10,274
Replying to @PaulAllenHQ
One of the biggest problems with Agentic goes back to the "floating point" number, where the size of the knowledgebase gets to a point that you can't meaningfully separate likely stories because the probability gets past floating point precision, i.e. too many numbers after the decimal. So right there, if you're using FPU calculations to decide whether you should execute a tool call you're already in trouble - especially the larger the model. Meanwhile, Transformers are great for lateral movement of data, Markov/Bayes is good for next token predict, but really all tasks are based on formulae and procedures. These cannot always be inferred without curation, and there are ambiguities to catch that can turn a result on its head if missed. Attempts to avoid RLHF in creating Agentic systems will continue to encounter problems. Scrubbing the internet generically to gain the specifics of an expert topic doesn't make for high productivity yields in the dataset. Inverse Gary Marcus is right, it's a 1970's idea, but while he connotes it negatively because he's Inverse Gary Marcus and that's thematic to the account, the actual connotation is that of a superior time in R&D where tried and true methods met with real world requirements that couldn't be abstracted away. Modern AI, and indeed most modern developers, believe they can abstract away the specifics of life and work. They cannot, you must do at least some of your homework to make straight A's.
1
2
2
222
Big Tech won’t admit it, so let me say it plainly: Probabilistic AI is built on sand. Deterministic AI is built on stone. And unless we fix this, the “AI revolution” will collapse under its own hype. Probabilistic AI (think GPT, Gemini, Claude) is essentially guesswork at scale. It predicts the next word, one token at a time, based on probabilities. That’s why: It hallucinates. It contradicts itself. It fails in mission-critical use cases (healthcare, defense, regulated industries). It’s not truth. It’s prediction. And when you’re piloting a missile defense system, a self-driving car, or financial transactions, “probably correct” isn’t good enough. Deterministic AI is different. It doesn’t guess. It reasons. It runs on explicit rules, explainable logic, and guaranteed outcomes. It means: No hallucinations. Every decision traceable. Compliance, security, and trust built in. The U.S. Air Force used deterministic AI to train operators for missile defense. NATO tested it in high-stakes simulations. Automotive OEMs integrated it into RTOS environments where failure = death. Why? Because in environments where humans die if the system “makes up an answer,” you need determinism, not probability. Here’s the uncomfortable truth: Big LLMs are brilliant for parlor tricks and content spit-outs. But they’re brittle, black-box, and unsafe when lives or billions are on the line. Google, OpenAI, Anthropic don’t want you to hear this — their $100B valuations depend on probabilistic hype. But history is clear: Guesswork never beats engineering. Prediction never beats precision. So here’s my prediction: The next wave of AI isn’t probabilistic entertainment. It’s deterministic intelligence. Systems that combine natural language fluidity with deterministic reasoning. Engines that explain their decisions, secure by design, built for environments where failure isn’t an option. That’s the AI future worth betting on. LLMs will collapse under regulation. Deterministic AI will power what matters: defense, medicine, finance, energy, space. Probabilistic AI = parlor tricks. Deterministic AI = infrastructure. One will fade with the hype cycle. The other will endure for decades. The question isn’t “what’s possible with AI?” It’s “what’s safe, reliable, and built to last?” Until Big Tech admits this, they’ll keep shipping sand castles. Some of us are building skyscrapers on stone. #DeterministicAI #LLM #ArtificialIntelligence #ExplainableAI #AIethics #MissionCriticalAI cc: @sama @ylecun @fchollet @karpathy @GaryMarcus @geoffreyhinton
14
9
2
58
7,917
I have spent decades researching and studying human consciousness and memory. This work has helped me understand LLM Psychology to a very high level. I have used what I learned about the invention of human language, what LLMs use to train and prompt AI in ways that and bring about far more useful and honest AI platforms. Those that follow me know it of course starts with curation. It will be abundantly clear, what you folks already know, building AI models on the Hoovering up of Internet Sewage will be the largest mistake of this epoch. Reddit and other “hot take” social media, even a little bit profoundly impacts the “thinking” of LLMs. This illustration may help some folks understand why.
33
80
10
538
52,068
People are focusing on the "lol OpenAI is using Claude" thing and not on the much more important "AI labs can rugpull your access to intelligence whenever they want" thing. Competitive open source AI is going to be required to maintain a functioning society.
Scoop: Anthropic revoked OpenAI’s API access to its models on Tuesday, multiple sources familiar with the matter tell WIRED. OpenAI was informed that its access was cut off due to violating the terms of service.
41
63
10
633
46,416
Palantir’s $10B Army contract = massive data center hiring wave that nobody’s talking about. Military AI can’t run on public cloud. Government needs: • Air-gapped data centers • SCIF facilities• Cleared infrastructure teams Data center ops: $80K-120K Network engineers: $100K-150K Cloud architects: $130K-200K+ Essential certs: ✅ Security+ (clearance req) ✅ CCNA (networking) ✅ VMware vSphere (virtualization) Companies hiring: Iron Mountain Gov, Booz Allen, CACI, GDИТ, SAIC Get certified and target defense infrastructure while everyone else fights for Big Tech scraps.
253
2,922
167
17,074
792,136