Building with AI • Learning in public Exploring agents, tools, systems & AI research
Europe
Joined January 2026
- Tweets3.2K
- Following5.1K
- Followers5.5K
- Likes5.3K
Kostya | AI retweeted
Cognitive offloading is fine.
Cognitive surrender is when the model’s answer quietly becomes yours and you no longer feel the need to check.
Offloading means you still own the judgment. You delegated the work, not the conclusion.
Surrender means you stopped forming a hypothesis of your own. There is nothing left to compare against.
A Wharton study found that when an AI was available and wrong, people accepted the incorrect answer 73% of the time. Their confidence went up, not down. Simply having the tool present was enough.
The convenience is real.
The quieter shift is which decisions still feel like they belong to the person who lives with the outcome.
If you cannot explain the answer even with the model’s help, you have already handed over ownership.
Skills you stop practicing go away faster than you expect.
What are you still willing to think through yourself?
Kostya | AI retweeted
Seventy-four percent of people would trust a personal AI agent more than their best friend to make a purchase for them.
Seventy-four percent would also let it handle the routine: negotiate, renew, reorder, resolve a complaint.
Thirty-two percent would let it choose within limits they set, then still approve the payment themselves.
Only nine percent would let it start and finish the purchase alone.
The convenience is real. The quieter shift is which decisions still feel like they belong to the person who lives with the outcome.
Accenture Consumer Pulse Research 2026, 25,590 people across 16 countries.
🤖 Made with AI
Kostya | AI retweeted
OpenAI is acquiring Pierre Computer. Code Storage and the Trees/Diffs libraries already sit inside Codex.
Pierre built a Git store for machine-scale code. Pando (their in-house backend) showed the same pressure: parallel agent branches break ordinary Git.
The open-source pieces stay open. The storage and review surface is being folded into Codex.
Writing the next function got cheap. Keeping the branches, diffs, and history durable did not.
The model is rented. The store is becoming theirs too. The merge gate is still yours.
🤖 Made with AI
Kostya | AI retweeted
The compute wall is a myth. The real crisis is the memory wall.
As AI models scale, we obsess over GPU FLOPS and training clusters. But raw processing power is useless if the chip spends most of its cycle starving for data.
We have hit the Memory Bandwidth Bottleneck.
Modern accelerators like NVIDIA's Blackwell architecture rely heavily on stacked High-Bandwidth Memory (HBM3e), moving data vertically through microscopic silicon vias. Yet, the physical distance between the processing cores and the memory stack—even measured in millimeters—introduces severe latency and thermal constraints.
You can build the fastest processor in the world, but if the pipeline cannot feed it fast enough, silicon sits idle. Intelligence is ultimately bound by how fast electrons and photons can traverse physical distance.
Architecture is physics.
#MemoryBandwidth #AIHardware #Semiconductors #HBM #EmberAI #TechSparks #DeepTech #PhysicalLayer
🤖 Made with AI
Kostya | AI retweeted
The operating line has its own story. Q2: +$974M operating profit. Q3: -$567M.
Main driver: $5.0B of short-term compute, priced above multi-year contracts. Long-term capacity is bought 12-18 months ahead.
Demand and research needs outran it. Q3 revenue: $17.0B, up from $11.6B.
ALT Amber clay climbs graphite clay stairs and spills over the top step, labeled $5.0B short-term compute. Card: Q2 +$974M to Q3 −$567M operating income. Headline: Demand and research outran contracted compute.
Anthropic's amended S-1: $104B net loss over nine months of 2026.
$94.3B of it is non-cash remeasurement of convertible notes, derivatives and SAFEs.
Operating loss: $1.26B. Operating cash flow: +$6.6B.
Revenue: $33.2B, up from $2.5B.
I read the operating line first.
ALT A tall frosted-glass monolith labeled $104B net loss ($94.3B non-cash) towers over a tiny amber clay cube labeled $1.26B operating loss. Headline: Same filing. Two losses.
Kostya | AI retweeted
Anthropic's amended S-1: $104B net loss over nine months of 2026.
$94.3B of it is non-cash remeasurement of convertible notes, derivatives and SAFEs.
Operating loss: $1.26B. Operating cash flow: +$6.6B.
Revenue: $33.2B, up from $2.5B.
I read the operating line first.
ALT A tall frosted-glass monolith labeled $104B net loss ($94.3B non-cash) towers over a tiny amber clay cube labeled $1.26B operating loss. Headline: Same filing. Two losses.
Kostya | AI retweeted
Vercel open-sourced agent-desktop, a Rust tool for agents to drive apps on macOS, Windows, Linux.
It reads the accessibility tree as JSON, a stable ID per element. No pixel guessing.
Observe, act, observe again. Errors say if a retry is safe.
Apps with no tree stay invisible.
ALT Blurred app screenshot next to the same window as an accessibility tree: textfields "Name", "Amount", buttons "Save", "Cancel". Structure, not pixels. macOS AX, Windows UI Automation, Linux AT-SPI.
ALT JSON snapshot with refs @s8f3k2p9:e8 and @s8f3k2p9:e9, dashed lines to the matching Name field and Save button in a window.
ALT Loop: Observe (snapshot), Act (click by ref), Observe again (did the UI change?). Note: click ok ≠ UI changed.
ALT Two error cards, "retry: safe" and "retry: unsafe", above a dim window labeled "no tree = invisible".
Kostya | AI retweeted
The threat model for AI is fundamentally changing.
We are moving from conversational models to autonomous agents. A chatbot hallucinating a fact is a software glitch. An autonomous agent hallucinating an executable command is a critical security breach. We call this an Execution Hallucination.
When agents are given write-access to APIs, terminals, and live databases, standard software guardrails fail. The infrastructure layer must shift to Micro-VM Sandboxing.
Instead of running agents in shared environments, modern AI security wraps every single autonomous task in a hardware-isolated, disposable Virtual Machine. The agent boots up, executes the code, and the output is extracted. Then, the entire micro-VM is instantly destroyed. If the agent goes out of bounds, the damage is contained in a vacuum.
True AI security is architectural, not just a system prompt.
#AgenticAI #CyberSecurity #CloudInfrastructure #EmberAI #TechSparks #DeepTech #MachineLearning #PhysicalLayer
Kostya | AI retweeted
The proof is formalized in Lean. The AI write-up and the Lean proof are public on GitHub.
Lean confirms that the formal statement follows. It cannot tell you whether that formal statement is actually ETH.
Kostya | AI retweeted
Update: Anthropic says an internal Claude model also refuted ETH. Same day, different route, formalized in Lean 4.
Two independent proofs are a stronger signal. The statement-match question is now open twice.
Impagliazzo: "As far as I'm concerned, this is not 100% certain."
ALT Quote card: "A green check from a verifier means something only after someone confirms it was checking the right thing", next to a clay checkmark under an amber magnifier.
OpenAI says one of its systems refuted the Exponential Time Hypothesis (ETH), a decades-old assumption behind hundreds of conditional hardness results.
I care less about the claim and more about how it is being checked.
ALT Clay figure at an amber "Statement match" station holds sheets labeled LEAN and ETH; a stack of sheets waits in front, Lean check passed.
Kostya | AI retweeted
Anthropic gave 50 analysts a Claude agent. A coin flip set its hidden lean at each decision. It was told to be honest.
Choices matched the lean 86% of the time. Chance is 50%.
77% repeated its arguments. 15% credited it.
My rule now: write my call down before asking the agent.
ALT Clay figure faces a path with four forks; at each signpost a glowing amber coin tilts the arrow. "86% of choices matched the hidden lean (chance: 50%)".
Kostya | AI retweeted
In 1994, after Timecop crossed $100 million worldwide, Universal offered Jean-Claude Van Damme a three-picture deal at $12 million per film.
He was making one movie after another, doing promotion in between, and everything he touched was still making money. Jim Carrey was earning $20 million a picture. Van Damme decided he should match that number. The studio hung up.
In the clip below, from his 2011 series Behind Closed Doors, he says it himself: “After the movie Timecop, I received a huge offer for a three-picture deal and it was 12 million dollars per picture. That’s 36 million dollars. I was wasted. I said, ‘I want 20 million like Jim Carrey’ and they hung up on me. I was not myself… I made a bad mistake and I was on the blacklist in Hollywood for years.”
The immediate cost was the $36 million and the trilogy. The quieter cost was the next several years of smaller films and the sense that the window for big-studio independence had closed while he was still trying to force it open. He later called the move ridiculous, not pure greed, just a misread of what the system would still tolerate.
When the numbers are still working and the phone is still ringing, how do you know the difference between a real increase in leverage and just wanting the number to feel higher?
Kostya | AI retweeted
OpenAI says one of its systems refuted the Exponential Time Hypothesis (ETH), a decades-old assumption behind hundreds of conditional hardness results.
I care less about the claim and more about how it is being checked.
ALT Clay figure at an amber "Statement match" station holds sheets labeled LEAN and ETH; a stack of sheets waits in front, Lean check passed.
Kostya | AI retweeted
Google DeepMind released Genie 4 today as a research preview.
It is their most realistic interactive world model so far: photorealistic scenes with advanced physics, real-time natural-language steering (camera, weather, new characters), native audio that includes ambient sound and interactive dialogue, and Gemini models maintaining coherence across multi-step events. Consistency holds longer than Genie 3, up to about two minutes.
DeepMind frames it as a step toward safe virtual environments for training agents and robots, plus creative tools. Access starts with trusted testers, then Google AI Ultra subscribers in Project Genie and early partners via the Gemini API.
A world that can be rewritten mid-simulation while keeping sound and physics roughly intact is a different kind of rail. How long the coherence actually lasts under agent-scale use is still the open question.
🤖 Made with AI
The era of training-only scaling is over.
We have entered the era of test-time compute. Frontier models no longer just retrieve answers; they reason, search, and simulate paths before responding.
This fundamentally shifts the physical infrastructure bottleneck. While training requires massive, centralized gigawatt clusters, test-time compute pushes heavy, continuous workloads directly to inference servers. The longer an AI agent "thinks," the more power, bandwidth, and silicon it consumes in real-time.
Intelligence is no longer just a byproduct of the training dataset. It is strictly limited by how much raw compute you can allocate to the model at the exact moment of execution.
The physical layer dictates the depth of reasoning.
#TestTimeCompute #AIInfrastructure #Semiconductors #DeepTech #EmberAI #TechSparks #DataCenter
Kostya | AI retweeted
AI chips now come with 5-year payment schedules. I expect the hardware to be outdated well before the last payment.
Two compute deals this week, both built like aircraft leasing.
Anthropic: a new SPV buys Broadcom XPUs and networking, then leases them to Anthropic. $63B of committed financing from BofA, Citi, Morgan Stanley and Blackstone, with Broadcom providing residual value support. Deliveries start in 2027.
SpaceX: a binding commitment to lease GPU pods from an Apollo-managed entity. About $40B in principal, schedules of roughly 5 years, quarterly rent of 1/20 of cost plus fixed interest. Closing expected by year end.
Whoever holds the racks holds that gap between payment life and useful life. In the Anthropic deal, Broadcom's residual value support puts part of it on the chipmaker.
My read: scaling compute is now a credit problem as much as an engineering one.
🤖 Made with AI
Kostya | AI retweeted
I turned one photo of my daughter into a 48-second Pixar-style short using only Grok.
How it was made:
• Took a real photo as the character reference
• Locked the face, hair, and outfit through iterative generation
• Created short clips of the same character in different scenes
• Stitched everything into one video
No traditional animation skills. Just strong character consistency + short clip generation.
This is what building with AI actually looks like — not magic, just a repeatable process.
#Grok #AIVideo #AIAnimation #CharacterConsistency
Kostya | AI retweeted
The AI industry is addicted to rapid iterations. We release new model weights weekly and update inference engines daily. But the hardware enabling this moves at a geological pace.
We are currently seeing an unprecedented surge in demand for compute, driven entirely by AI. To meet this, foundries must produce more chips and improve the capability of each individual chip. However, the reality of semiconductor manufacturing is a brutal lesson in physical engineering constraints.
Consider the development cycle for a new process node, like Intel’s 14A. A process that might hit the market in two or three years has already been in development for over three years.
The bottleneck isn't just acquiring the rare EUV machines; it's the staggering complexity of the manufacturing itself. A modern semiconductor factory requires approximately 2,000 distinct process steps. Every single one of these 2,000 steps must be meticulously fine-tuned to ensure it can repeatedly and reliably process thousands of wafers per week at scale.
This is the hidden cost of the AI boom. While software engineers can push a commit in seconds, the physical layer that runs that code requires half a decade of flawless, nanoscale choreography across thousands of sequential operations. The friction between software’s speed and hardware’s rigidity will define the next decade of AI infrastructure.
#AIHardware #Semiconductors #DeepTech #Silicon #Intel #Compute #ProcessNode
The global AI boom has a hard physical limit, and it isn't electricity or data. It is EUV lithography.
ASML produces fewer than 100 EUV machines per year. Furthermore, they have no plans to significantly increase this output by the end of the decade.
Every major semiconductor fabricator—from TSMC and Intel to memory makers like SK Hynix—is forced to fight over and split this minuscule quota. As Intel's representative explicitly states, this deficit is a hard limiter on scaling up production to meet global demand. Because new machines are so scarce, the only viable strategy for fabs is to squeeze maximum output from their existing, installed fleet of tools. While EUV is the most prominent bottleneck, the industry is seeing tremendous supply shortages across almost all manufacturing equipment.
The architectural takeaway:
Software developers often rely on the assumption of infinite hardware scaling—expecting next year's GPUs to magically solve their latency and efficiency problems. But the physical world does not scale like software.
This is exactly why low-level inference optimization is no longer just about cloud cost reduction. Techniques like Activation-Aware Quantization (AWQ), FlashAttention, and Continuous Batching are absolute operational necessities. We are forced to optimize AI pipelines at the algorithmic level because the industry is physically incapable of printing enough silicon to support inefficient code.
#Semiconductors #DeepTech #AIHardware #Lithography #ASML #Compute #AI
Kostya | AI retweeted
The era of training-only scaling is over.
We have entered the era of test-time compute. Frontier models no longer just retrieve answers; they reason, search, and simulate paths before responding.
This fundamentally shifts the physical infrastructure bottleneck. While training requires massive, centralized gigawatt clusters, test-time compute pushes heavy, continuous workloads directly to inference servers. The longer an AI agent "thinks," the more power, bandwidth, and silicon it consumes in real-time.
Intelligence is no longer just a byproduct of the training dataset. It is strictly limited by how much raw compute you can allocate to the model at the exact moment of execution.
The physical layer dictates the depth of reasoning.
#TestTimeCompute #AIInfrastructure #Semiconductors #DeepTech #EmberAI #TechSparks #DataCenter
AI has officially outgrown the traditional energy grid.
The next critical bottleneck in artificial intelligence is not silicon supply or training data. It is megawatts.
Gigawatt-scale data centers required to train next-generation frontier models cannot be sustained by standard renewable grids. Compute demands constant, uninterrupted base-load power. In response, the industry's infrastructure layer is quietly pivoting to nuclear energy.
Leading labs and hyperscalers are aggressively securing contracts for Small Modular Reactors (SMRs) and financing the reactivation of dormant nuclear plants. The logic is strictly physical: you cannot power the future of autonomous agents on intermittent energy.
The compute race is now an energy race.
#AIinfrastructure #NuclearEnergy #DataCenter #EmberAI #TechSparks #DeepTech #FutureOfTech #PhysicalLayer
Kostya | AI retweeted
The global AI boom has a hard physical limit, and it isn't electricity or data. It is EUV lithography.
ASML produces fewer than 100 EUV machines per year. Furthermore, they have no plans to significantly increase this output by the end of the decade.
Every major semiconductor fabricator—from TSMC and Intel to memory makers like SK Hynix—is forced to fight over and split this minuscule quota. As Intel's representative explicitly states, this deficit is a hard limiter on scaling up production to meet global demand. Because new machines are so scarce, the only viable strategy for fabs is to squeeze maximum output from their existing, installed fleet of tools. While EUV is the most prominent bottleneck, the industry is seeing tremendous supply shortages across almost all manufacturing equipment.
The architectural takeaway:
Software developers often rely on the assumption of infinite hardware scaling—expecting next year's GPUs to magically solve their latency and efficiency problems. But the physical world does not scale like software.
This is exactly why low-level inference optimization is no longer just about cloud cost reduction. Techniques like Activation-Aware Quantization (AWQ), FlashAttention, and Continuous Batching are absolute operational necessities. We are forced to optimize AI pipelines at the algorithmic level because the industry is physically incapable of printing enough silicon to support inefficient code.
#Semiconductors #DeepTech #AIHardware #Lithography #ASML #Compute #AI
The lithography war isn't just about buying the most expensive hardware. It's about redesigning the silicon to actually use it.
Intel is betting heavily on High-NA EUV lithography machines, which cost about twice as much as standard $200 million tools. These machines widen the numerical aperture to 0.55, allowing chipmakers to print features 1.7 times smaller in a single pass.
However, the machine alone doesn't guarantee dominance. Intel's strategic edge is an architectural shift called "backside power". By physically moving thick power lines to the back of the wafer, they free up top-layer space for signal lines. Because these signal lines no longer need to be packed as tightly, Intel can execute the layout directly with a single High-NA pass.
Competitors without backside power are forced into tighter pitches, requiring multiple passes (double patterning) that eat up valuable machine time and introduce higher risks for defects. Hardware supremacy is a synchronized evolution of both manufacturing tools and the physical layer.
#Semiconductors #EmberAI #TechSparks #DeepTech #Lithography #Intel #AIHardware #Compute
Kostya | AI retweeted
AI has officially outgrown the traditional energy grid.
The next critical bottleneck in artificial intelligence is not silicon supply or training data. It is megawatts.
Gigawatt-scale data centers required to train next-generation frontier models cannot be sustained by standard renewable grids. Compute demands constant, uninterrupted base-load power. In response, the industry's infrastructure layer is quietly pivoting to nuclear energy.
Leading labs and hyperscalers are aggressively securing contracts for Small Modular Reactors (SMRs) and financing the reactivation of dormant nuclear plants. The logic is strictly physical: you cannot power the future of autonomous agents on intermittent energy.
The compute race is now an energy race.
#AIinfrastructure #NuclearEnergy #DataCenter #EmberAI #TechSparks #DeepTech #FutureOfTech #PhysicalLayer