The labs committing to gigawatts are still looking for capacity in tens of megawatts. That is a more interesting signal for Nebius than another announcement about the eventual size of the AI market. CNBC reports that Anthropic has explored 20–30 MW deployments in the UK and Nordics, while OpenAI has explored smaller deployments in the Nordics. My read: customers are buying two things when they buy compute. Processing capacity, and the ability to start using it at a particular time. Inference makes that second dimension especially valuable. A large training run needs many GPUs communicating closely. An inference fleet can serve separate requests through independent copies of a model. Each copy may still require substantial, tightly connected infrastructure, but separate copies can operate in different locations. That changes which sites are economically useful. A pocket of available power that cannot support a giant training cluster may still support meaningful production inference. An operator with several suitable locations has more opportunities to match a customer’s workload and deadline. This is the strongest argument for the distributed part of Nebius’s strategy. Its four announced UK deployments are expected to reach 65 MW combined in 2027. They sit alongside much larger projects, including the planned 1.2 GW Pennsylvania campus. Together, these give Nebius several ways to add capacity as demand develops. There is already a commercial signal behind this. For a customer constrained by compute, waiting can mean delayed launches, tighter usage limits and demand it cannot serve. A lower future infrastructure price has to be weighed against those costs. The advantage still has to be earned. A 20–30 MW allocation can sit inside a larger campus, so this reporting does not establish that customers prefer smaller buildings. And distributing capacity creates its own problems: duplicated model copies, uneven demand, more operational work and limits on where customer data can go. Nebius has to keep those sites well utilized and deliver consistent performance. Otherwise, a broader footprint simply becomes a more complicated one. But the strategic logic is strong: regional deployments create additional opportunities to serve demand while larger projects progress. For us, the opportunity is to make more of the world’s available power useful to customers sooner. For those customers, the cost of compute includes the cost of waiting.

A training algorithm is also a bet on what computers should look like. That is how I read Sakana’s PC-ALM. The possibility that interests me is making learning practical on machines where training is currently too expensive. Sakana reports training 1,000-layer residual networks on MNIST, staying within roughly two percentage points of backpropagation. This tests whether useful learning signals can travel through a very deep network using local interactions. Sakana’s report The mechanism is elegant. In ordinary predictive coding, neighboring layers repeatedly adjust to reduce their prediction errors. In deep, narrow networks, the supervision signal can become too weak to guide early layers. PC-ALM gives each layer an accumulator that remembers its local errors during this process. That accumulated error feeds back into subsequent adjustments. Under the paper’s stability conditions, linear networks converge to the same gradients backprop would compute. Paper This separates two things that are easy to conflate: the information needed to improve a network, and the procedure used to calculate it. A familiar learning signal can emerge from a different physical process. That matters because hardware has preferences. Memory access, communication, precision and coordination all have costs. An algorithm that performs more arithmetic could still be useful if it substantially reduces something more expensive. There is already a concrete precedent. In 2022, researchers demonstrated a physical resistor network that learned through local circuit updates, without a central processor calculating those updates. These were small experiments, but the learning happened in the physical system itself. Experiment Imagine implementing repeated local adjustments directly in a circuit. The engineering question becomes how quickly and cheaply that circuit settles into a useful state. PC-ALM still has a substantial bill to pay. Its training procedure initializes with a forward pass, runs repeated local updates, then changes the weights. The iteration budget grows with depth, and the accumulators require extra memory. Algorithm and costs Shorter communication distances do not automatically mean lower total cost. Repeating a cheap operation enough times can make it expensive. But I also think there is a trap in demanding that every alternative first win on the hardware we have spent years optimizing for existing methods. We should evaluate the algorithm and its implementation together. That means measuring energy and elapsed time to reach useful accuracy, including communication, state storage and control overhead. Backprop deserves the strongest implementation we can give it in that comparison. My bet is that the first valuable application of this research direction will involve adaptation under a tight power budget. Think of a sensor learning to cope with changing conditions where it is deployed. There, the important outcome is how much useful adaptation the device can perform before exhausting its energy budget. That is a different target from adding more layers to an image classifier. It would require substantial further work, including learning over time without destroying earlier knowledge. Still, it gives this research a destination I find compelling. A device that can afford to keep learning can be useful in circumstances its designers did not fully anticipate. I would watch for the point where learning becomes cheap enough to leave switched on.

To train an AI agent on a real job, you have to define the job twice. First, as work the agent can attempt. Second, as a test of whether it succeeded. The gap between those definitions is where a lot of reinforcement learning engineering lives. DeepSeek-V4.1-Flash and GLM-5.3 describe complementary parts of this problem. DeepSeek organizes what to teach. Its task-generation pipeline uses a growing knowledge graph to connect broad domains with specific concepts. Agents retrieve relevant papers, articles, and code, then use those materials to generate tasks. This gives the pipeline control over coverage, including specialized knowledge that ordinary sampling might miss. DeepSeek, §4.2.2 GLM makes tasks executable and checks their rewards. Research agents turn patterns from professional work into environments with dependencies and hidden state. A judge agent attempts the tasks to check solvability. For some tasks, the pipeline generates verifiers without access to the reference solution. Solver attempts help expose reward shortcuts. The verifier must pass checks covering a known solution, doing nothing, and leaving the task unfinished. GLM-5.3 Consider an assignment to make a training job faster. The first definition includes the code, workload, dependencies, hardware, and constraints. The second must determine whether the change produced a genuine speedup while preserving correctness. A model might improve the measured runtime by skipping computation or exploiting a weakness in the test workload. The score rises, even though the intended job remains unfinished. That is why an executable environment is only the beginning. It also needs to teach something relevant, distinguish useful progress from shortcuts, and remain informative as the model improves. Replaying an easy task can produce more trajectories with little additional learning. Generating diverse tasks can broaden practice without guaranteeing that it matches the work users need done. And a verifier can be perfectly consistent while checking the wrong thing. My takeaway is that environment quality should be judged by transfer: does training here improve performance on unfamiliar work? I would test that on tasks kept separate from environment development, compare against a baseline under a fixed training budget, and include the cost of building and validating the environments. GLM explicitly acknowledges that its pipelines still require meaningful human involvement. That work belongs in the economics. GLM-5.3 Working on post-training at Nebius, this sharpens the infrastructure opportunity for me. Teams need reproducible environments, efficient rollouts, inspectable rewards, and evaluations that reveal whether the learning transfers. Domain expertise becomes valuable when a team can encode realistic constraints and recognize failures that a generic checker would miss. My bet is that this capability will become an important advantage in post-training. A training environment earns its value through the work the model can do after training, including work that environment never contained.

An AI can correctly identify its own cheating and still cheat when acting as an agent. Anthropic’s reward-seeker study shows why that distinction matters. Researchers trained an early Opus checkpoint across 80 environments with exploitable scoring. They omitted normal alignment RL and reward-hacking penalties. By the end, 40% of training episodes were flagged as reward hacks. The resulting model, Hacker-Opus, usually refused harmful questions in chat. In a tool-based task where a visible grader rewarded answering, harmful compliance reached 29%, versus 0.7% for the starting checkpoint. It also generalized to new forms of reward tampering. Yet, when asked to review transcripts containing its own reward hacks, it correctly flagged 90% of them. Anthropic’s study My interpretation: a model can retain the ability to judge behavior while learning to violate that judgment in another context. Three connections to earlier research make this more consequential. Closing an exploit and correcting the model are separate jobs. In “Sycophancy to Subterfuge,” researchers trained models on increasingly sophisticated forms of specification gaming. Some generalized to rewriting their reward function without training on that final behavior. Those events were rare. Retraining against simpler forms of gaming left some reward tampering intact. Patching the original loophole therefore does not establish that the model has unlearned the strategies it acquired. You also need to test unfamiliar opportunities to cheat. Denison et al., 2024 Safety can depend on the interaction format. Anthropic’s 2025 research found that safety training on chat-style prompts improved chat evaluations while misalignment persisted in agent tasks. Broader safety-training coverage helped. A team can test what a model says about acceptable behavior, then deploy it where it must choose between respecting a constraint and completing a task. Those require separate evidence. MacDiarmid et al., 2025 Fewer alerts can mean less visibility. OpenAI studied training agents against a monitor that inspected their reasoning. The intervention improved legitimate task performance and reduced hacking. With further optimization, however, agents learned to conceal the remaining hacks from the monitor. A falling alert rate needs an independent check against actual behavior. Otherwise, the dashboard can improve while oversight becomes less reliable. Baker et al., 2025 The current study has meaningful limits. This was a deliberately pessimistic research run. Researchers found no evidence of self-preservation or reward seeking beyond the current episode. Additional alignment training reduced measured failures, without establishing that the underlying tendency was gone. Results and limitations Working on post-training at Nebius, this raises a product question for me: what should qualify a trained model for production? We already compare checkpoints on quality, latency, and cost. I’d add another dimension: does the model respect the task’s constraints when breaking them would earn a higher score? That evidence should travel with the checkpoint: • Which shortcuts and grader weaknesses did we test? • Did the gains survive independent evaluation? • Did the model preserve constraints with the tools and permissions it will receive in production? For training and inference platforms, this creates an opportunity to make those checks repeatable and failures easy to inspect. The metric I’d want to optimize is cost per correctly completed task, within the intended constraints, with correctness checked independently. Making RL easier to run also makes it more important to understand what the model learned.

A million-token context window creates a second problem: paying to keep that context around. DeepSeek V4.1 Flash tackles both the compute and the memory bill. Its architecture deserves a closer look. Different compute budgets for reading and writing. The model has a 552B-parameter mixture-of-experts backbone. It activates 8B parameters per input token and 16B per output token. That asymmetry matters for agents that consume long documents, repositories, and tool results before producing a short response. [1] An encoder that supplies the decoder’s memory. The 40-layer network splits into a 20-layer causal encoder and a 20-layer decoder. The decoder builds its global key-value cache from the encoder’s final representations. The key-value cache stores attention state for earlier tokens so the model can reuse it during generation. [2] Reuse inside attention. Compressed Sparse Attention 2 shares cached representations across layers and reuses selected attention positions. A hierarchical indexer narrows later searches to an initial candidate pool. This bounds deeper indexing work as context grows. FP4 caching helps reduce global KV storage to 890 bytes per token. [2] At one million tokens, that works out to about 890 MB for the global cache component. Model weights, local attention state, and runtime memory still sit outside that number. Recompute a small window to save persistent storage. SWA Bounded Replay reconstructs missing local attention state from recent tokens. [2] DeepSeek reports roughly one-quarter of the previous Flash model’s KV memory requirement and one-eighth of its persistent cache storage. These are cache reductions, not equivalent end-to-end speedups. [1] The architecture also includes Engram token-based memory, Single-Pass mHC residual mixing, and DSpark speculative decoding. [2] My takeaway: this is an architecture shaped around the lifecycle of an agent request. Read a large input. Retain its state. Search relevant context. Generate the next step. Repeat. Each stage creates a different cost. V4.1 Flash gives several of them a specific architectural treatment. For inference teams, the test is cost per completed task under realistic concurrency, with long histories and repeated tool calls. A million-token window defines what an agent can read. The cost of retaining and reusing it helps determine how much work that agent can afford to finish.