In 2026, we hope to put NonBioS to sleep.
If you have used NonBioS for some time, you will realize that it starts forgetting key details after a few hours. It stops being useful. If you are working on a large software task, you need to reset and start a new chat. This is a fundamental constraint with NonBioS, and most agents in general. The anthropomorphic interpretation of this state is that NonBioS gets “tired.” Exactly like a human after a long day of work. Spot fixes, like caffeine for humans, will only get you so far. The real solution, in anthropomorphic terms, is that NonBioS needs to “sleep.” In 2026, we plan to be able to put NonBioS to sleep. Let me dig into what this means and what this will change.
The core reason that NonBioS starts forgetting is that its context fills up. The context is something like a short-term memory for agents. As NonBioS works through a large software task, it starts understanding pieces of the task. This “understanding” resides in its context. The key capability that we have engineered NonBioS with is aggressive context management. We implemented what we call “Strategic Forgetting” which prunes this context, keeping only what is important and discarding the rest.
However, this capability is not perfect. Over time NonBioS forgets the very first instructions itself, what the task it set out to do in the first place was. This degradation is fundamental, and reinforcing the objective is only transient. The key problem is that the context has become too information dense. Reinforcing the objective only serves NonBioS for some more time before it loses track once again. If you have used NonBioS for some time, this becomes apparent in its responses. To make further progress, we advise our users to start a new chat. This gives NonBioS a blank slate, but because all the work happens in the Linux sandbox, NonBioS is able to quickly gather context and restart its work.
But why can’t we just keep enlarging the context? The short answer is that current LLM architectures make it very expensive. More importantly, as the context becomes large, the ability of NonBioS to apply intelligence decreases. The larger the context, the more information it contains. The more information that NonBioS has to use a priori, the less effective intelligence it demonstrates. Expanding context is a zero-sum game, and very soon the only option is to reset and restart.
The long-term solution for this problem is to put NonBioS to sleep. When humans go to sleep, they clear their heads. But most importantly, they transfer whatever is there in short-term memory to their long-term memory. This does two things: it clears out short-term memory to tackle tasks the next day, and it stores the important learnings of the previous day to create continuity. This continuity is a result of continuous learning, for which sleep is the key enabler in humans. In exactly the same way, we hope to clear NonBioS’s context while transferring the learnings of that session into the model itself.
How might this work technically? The context engineering we already do with Strategic Forgetting provides a foundation. We are already identifying what information is important and what can be discarded. The next step is to take what we identify as important and, instead of merely preserving it in context, encode it into the model’s weights. The orchestration layer that currently manages context pruning would be extended to manage this transfer during “sleep” cycles. The mechanisms for this are still being developed, but the core insight is that our existing context engineering architecture points the way forward.
Continuous learning may prove to be one of the most important capabilities for AI as we head into 2026. I believe that continuous learning is likely a necessary condition for something approaching Artificial General Intelligence to emerge, and I suspect it might also be sufficient. We may be a few years away from finding out. At NonBioS, this is the year we start moving toward this goal. I suspect that a physical embodiment of a continuous learning system may exhibit properties that resemble sentience and consciousness, though these terms carry philosophical weight that makes precise claims difficult. This would give us something like the first Non-Biological Sentience, or NonBioS for short.
What does continuous learning capability mean in practice? If you are building a SaaS application through NonBioS today, you need to do it part by part. You first build the frontend, then the backend, then each infrastructure component: billing, authentication, multi-tenancy, dashboards, notifications, and so on. Each such feature takes a few chats to build. NonBioS starts on a feature, builds some parts, and then after a couple of hours loses the plot. You start a new chat, NonBioS picks up context, and then takes it ahead. Feature by feature, part by part, a full SaaS application comes together. The biggest ones in production right now on NonBioS are a hobby school management system and a classifieds portal. Each took about a month or two to build, with the user actively shepherding NonBioS through tasks.
Once continuous learning is built into NonBioS, this completely flips the game. When NonBioS gets tired after a couple of hours, we will put it to sleep. NonBioS will not be active for the next hour or so. Once NonBioS awakens, it will have a clean context but with an important difference: it will also know what it did in the previous session. You will not need to get NonBioS up to speed. NonBioS will know what the larger objective was, say building the classifieds portal, and which task it was on, say building the billing component. Once NonBioS awakens it will start to build again, seamlessly taking up where it left off and making progress. NonBioS will be able to build entire SaaS applications with the user checking in periodically or being interrupted for updates.
At this point, we don’t know how effective continuous learning will be in NonBioS, or how much it will cost. But we do have a roadmap for getting the first NonBioS “alive” within a continuous learning framework sometime this year. Over the next few years, we expect this to go down in cost and improve in efficacy, allowing entire software systems to be built in a largely autonomous manner.
I think the bigger discussion we need to have is what continuous learning will mean to our society and economy. In the context of software engineering, continuous learning has the potential to automate significant portions of engineering work, including tasks that currently require senior-level expertise. The exact percentage is hard to predict, but it could be substantial. The job displacement we saw in 2025 may be an early signal of larger shifts to come. And NonBioS is not the only lab working on continuous learning. If we don’t do it, someone else will, very soon.
Software jobs are not just another category of employment. They are among the most highly paid jobs in the world, and for good reason. Architecting good, reliable software requires extreme cognitive capability. This has been a scarce capability in humans. But once this is in the realm of automation, the implications extend beyond software. A system that can reliably automate a large fraction of software engineering roles might, with modifications, be applicable to many other knowledge work domains. The full scope is uncertain, but the direction seems clear.
I believe we as a society are not fully prepared for such a transformation. The competitive pressure in a capitalist society does not easily allow for the kind of deliberate, thoughtful transition that a change of this magnitude requires. Different stakeholders, including governments, companies, workers, and researchers, will need to coordinate in ways that do not come naturally in a competitive landscape. What might this look like? It could involve investment in retraining and education, exploration of new economic models, and serious conversation about how the gains from automation are distributed. These are hard problems, and there are no easy answers.
I believe mankind would benefit from a pause to think carefully before we move ahead.
Hopefully we can take one in 2026, before we put NonBioS to sleep.