AI Thesis · 01

Continual Learning

Continual learning ranks among the hardest unsolved problems in AI research. It is a primary bottleneck on the path to recursive self improvement.

Jason Rubenstein

Continual learning is the ability of a model to adapt to new, changing data while it keeps what it already knows. Real world data shifts on its own schedule. A customer service tool has to learn new company policy. A recommendation system has to track a user's changing taste. Without continual learning, a model loses accuracy as its environment moves past it.

Leading startups in this space raised more than $150 million in the last three months. The biggest beneficiaries will be companies that own unique inference data and workflows. Cursor is one proof point, turning its own inference data into training signal for real time reinforcement learning.

Building this infrastructure takes thousands of GPUs, interconnect systems, enough memory to hold weights and computation, and a team of highly paid researchers. Most workflow owners hold the data but not that infrastructure, which opens room for standalone continual learning startups to fill the gap. Prime Intellect, Engram, and Trajectory illustrate three ways to help workflow owners build their own models on an open weight base.

How training works

Billions of dials, frozen at ship time

Every weight in a model is a dial set during training. The model sees an example, guesses, and nudges the relevant dials toward a better guess. Once training ends, the dials lock. Drag the control to see what changes when a model keeps learning after it ships.

Frozen model Continual learning

Training nudges billions of weights until the model's guesses match reality. When training stops, every dial locks in place. The model answers only from what it learned before that moment, no matter how the world changes after.

Continual learning keeps nudging the dials after ship day. New data updates the same weights, so the model absorbs fresh facts and feedback without a full retrain. New nudges can also disturb the settings that held old knowledge, so the model forgets what it knew.

Turning the dials to learn something new disturbs the settings that held old knowledge. The model can forget what it knew.

From the thesis
Three approaches

Three ways to keep a model current

Startups attacking continual learning split into three camps. Each pushes the cost of staying current to a different place in the stack.

Parametric modification changes a model's own weights after it ships. The update lives in the weights, so the model carries the new behavior everywhere it runs, with no reminder needed at query time.

The trade-off is cost. Editing weights without disturbing old knowledge is slow, expensive, and the core problem every company in this camp is solving.

Standalone companies here face a structural choice. Build on an open weight model such as Kimi, DeepSeek, or GLM, and value tends to accrue back to the open source lab. Partner closely with a closed weight lab such as OpenAI, Anthropic, or DeepMind, and the more likely outcome is an eventual acquihire.

Where it lands

Cost lands in
Training compute
Training vs inference
Training heavy
Maturity
Early, venture scale race
Example companies
Prime Intellect, Engram, Trajectory

These startups treat memory as retrieval, not as a weight update. The model's weights stay frozen. A separate store of working memory updates instead, and the model pulls from it at query time.

Memory complements continual learning rather than replacing it. Differentiation in this space has proven hard to sustain. Most players get mixed feedback from users, and many have shifted their pitch toward caching and cost savings instead of a better product experience.

Foundation model providers see user memory as a lever for switching costs. That makes this category harder for a standalone startup to defend.

Where it lands

Cost lands in
Inference, per query
Training vs inference
Inference heavy
Maturity
Crowded, weak differentiation
Example companies
Mem0, Letta, Zep

This camp updates the physical processor. Instead of shuttling data between separate compute and memory, these chips store a weight and compute with it in the same physical spot, closer to how a brain works.2

The rationale starts with biology. A human brain learns for life, holds decades of memory, and runs on about twenty watts. A frontier model burns far more energy and still forgets easily. Proponents argue the chip itself, not only the algorithm, is part of the problem.

Unconventional AI looks the most relevant name in this group. Founder Naveen Rao co-founded MosaicML, later acquired by Databricks. His claim is that an analog chip can behave like the math a model needs to run, rather than simulate that math step by step through transistors.

Where it lands

Cost lands in
Chip fabrication, R&D
Training vs inference
Neither yet, cost sits in design
Maturity
Early, research and prototype
Example companies
Unconventional AI, Extropic
Common misconceptions

Two ideas people get wrong

They sit at opposite ends of cost and speed. A weight update is a training run, and its cost scales with how often and how many models get updated. Retrieval is fast and adds little beyond the tokens it pulls in.

Even where a weight update makes sense, most systems still want a retrieval layer too. Facts pulled from an external store can be cited, audited, and edited cheaply. Facts baked into weights cannot.

A model that cannot take in new information will never improve itself, so continual learning is a precondition. It is not the hard part.

The harder capability is judgment. That means choosing what to learn, what to discard, and where to push next. A model can update its weights all day and still choose bad problems. Continual learning clears one barrier to recursive self improvement. It does not deliver it in full.

Compute implications

Where the cost lands

The two leading approaches push spending to different parts of the stack.

Parametric pushes cost into training

Training
Inference

Every weight update is a training run. The more often a model relearns, and the more models a provider keeps current, the higher the standing compute bill. The spending flows to compute providers and to whoever owns the update pipeline.

Memory pushes cost into inference

Training
Inference

Freezing the weights avoids the retraining bill, but the cost returns at run time. Every call retrieves context and feeds it back through the model, burning tokens per request. The spending lands on inference providers and on memory tooling instead.

How to read this. The bars show where each approach concentrates spending, not a measured split. Neither source document gives an exact ratio between training and inference cost.

If every startup in this space converges on one approach, the leading labs capture the value. Everyone else exits through acquihire.

From the thesis
A leading indicator

Recursive self improvement, in one timeline

Continual learning is necessary for a system to improve itself. It is not sufficient on its own. Anthropic's own numbers, from its June 2026 piece "When AI Builds Itself," show how fast the easier part is moving.

Before 2025

Claude writes almost none of Anthropic's code

Claude's share of code merged into Anthropic's own codebase sits in the low single digits.

November 2025

Best model beats the human call about half the time

On a test where models pick the next step in a real experiment, Anthropic's best model matches or beats the human choice 51 percent of the time.

April 2026

That figure climbs to 64 percent

Four months later, the same benchmark puts the model's win rate against the human choice at 64 percent.

May 2026

Claude writes most of the code

More than 80 percent of the code merged into Anthropic's codebase is authored by Claude, and the typical engineer ships eight times as much code per day as in 2024.

Nov 2025
51%
Apr 2026
64%

How to read this. Anthropic tested moments where a researcher's next step had room to improve, then had a separate model judge whether the AI's suggested step or the human's actual choice worked out better. The figures above show how often the AI's suggestion won that judgment.

None of this means recursive self improvement has arrived. Anthropic says plainly that it has not, and that it is not inevitable. The capability still missing is judgment, the same gap flagged above. Continual learning is worth tracking as a leading indicator toward that gap closing, not as proof that it has closed.

Company guide

Who is building this

Grouped by approach. This is not the first generation of continuous training platforms. Predibase and Lamini pursued continuous fine-tuning years earlier, and each was acquihired once adoption stalled.1 The trend toward open source, post-trained models has widened the opening for a similar pitch to land this time. Cards marked open-weight focus help workflow owners build their own models on an open weight base.

Parametric modification

Update the weights directly

Verifiable continuous learning platform that intercepts agentic workflow failures and user feedback loops.

An agent keeps booking the wrong meeting room. Relai catches the failure and the user correction, verifies the fix, then folds it into the model.

Founder: Soheil Feizi

Runs a full lifecycle platform, Catalyst, that captures production traffic and eval failures, turns them into training data, then fine-tunes and hosts task-specific custom models.

A team runs a cheap custom model to tag support emails. As new ticket types show up, fresh production data retrains the model to keep pace.

Founders: Francesco Virga, Ibrahim Ahmed, Amarjot Singh

Develops deep-tech database infrastructure built to feed continuous streaming loops directly into model weights.

A vision model misses a new product on the shelf. New labeled images stream from the database into the next training pass.

Founder: Davit Buniatyan

Trains custom specialist models for a customer's domain by RL post-training an open model against that customer's data, evaluations, and expert judgment.

A 7B open model RL post-trained for JEE math went from 54.7% to 90.2% accuracy, beating o4-mini and Gemini 2.5 Flash.

Founders: Sachin Dharashivkar, Rohith Pesala

Non-parametric memory

Frozen weights, updated working memory

Acts as a continuous learning engine by programmatically generating prompt fixes that repair behavioral drift instantly, with no weight change.

After a model change, an agent starts greeting users too casually. Lemma spots the drift and writes a prompt patch that corrects it.

Founders: Jerry Zhang, Cole Gawin

Extracts, enriches, and structures dialogue histories continuously into an accessible graph layer that persists across sessions.

A user mentions a peanut allergy in chat. Zep pulls that fact into a graph so the assistant recalls it in later sessions.

Founder: Daniel Chalef

Creates an intelligent, self-improving long-term memory layer for AI applications that adapts based on user interactions.

A user says they prefer aisle seats. Mem0 stores the preference and surfaces it the next time they book a flight.

Founders: Deshraj Yadav, Taranjeet Singh

Enables agents to manage, compress, and read-write to their own persistent storage subsystems.

An agent runs low on context mid task. Letta lets it write notes to its own store and read them back later, like a working notebook.

Founders: Charles Packer, Sarah Wooders

Open-source AI engineering and evaluation tracing loops that let development teams monitor agent drift.

A team notices answers getting worse. Langfuse traces each agent step so they can see where the drift began and fix the prompt or tool.

Founders: Marc Klingen, Max Deichmann, Clemens Rawert

Neuromorphic hardware

Update the physical processor

Rethinking the foundations of the computer to optimize energy efficiency for AI, building silicon circuits with non-linear dynamics similar to biological intelligence.

Founders: Naveen Rao, Michael Carbin, Sara Achour, MeeLan Lee

Its biological arrays lean on natural cell plasticity to continuously learn and process workloads at near-zero power.

Founder: Hon Weng Chong

Computing processors that integrate synthetic biological cells onto physical chips rather than using traditional transistor circuits.

Founder: Oshiorenoya Agabi

Develops ultra-low-power neuromorphic processors built for sensor data and edge computing applications.

Founder: Sumeet Kumar

Builds thermodynamic sampling units, pbits, that use thermal noise for probabilistic AI at a fraction of the energy.

Founders: Guillaume Verdon, Trevor McCourt

Evaluation framework

Questions for a founding team

Founders of the earlier fine-tuning wave share traits worth studying today.3 Two sets of questions help evaluate a team's technical readiness.

On team composition

  • Who owns the research, and who owns the engineering? Are they the same person?
  • Who on the team has built large-scale training or high-traffic serving systems before, and where?
  • Has anyone on the team shipped something to real users at scale, or is this the first time?
  • What gap in the team are you hiring for next?

On priority and readiness

  • Data scarcity. Are you generating high-quality, workflow-specific data that would be hard for others to replicate?
  • Model agnosticism risk. If your product sits on top of leading foundation models with no proprietary learning loop, does that create real competitive risk?
  • Ability to act on the data. Are you capitalized well enough to attract strong AI engineers and fund the compute this takes, or does a distributed RL partner make more sense?
  • Stage and timing. Is the company mature enough that this should be a priority now, rather than a capability worth revisiting later?
Market view

Where the opportunity stands

Prime Intellect is one notable infrastructure layer for neolabs such as Ricursive and Arcee, providing compute, managed training, inference, and research tooling. Reinforcement learning workloads drive most of its revenue, post-training is its fastest growing line, and on-demand GPU rental stays under ten percent of revenue. The team treats continual learning as a north star, and it has driven production wins including Cursor's Tab model and now Composer.

A founder's published research can signal a company's likely direction. Engram co-founder Jessy Lin published a relevant paper during her PhD at Berkeley. Work like that helps show whether an approach is differentiated and whether it can scale past a proof of concept.

If every startup in this space converges on one approach, the leading labs, Anthropic, OpenAI, Google, capture most of the value, and outside teams tend to exit through acquihire. The opportunities that last are adaptive systems built around a specific workflow. Think personalization, real time threat detection, or recommendation systems that shift with behavior. Engram's work with Notion and Harvey is an early version of this pattern.

Two revenue paths look credible. One is enterprise specific models updated on a rolling basis, so each company runs a model trained on its own workers and workflows. The other is a foundation model of its own for any team that builds real user traction, priced at a premium today. Further conversations with founders will test these views.

Footnotes

  1. Predibase and Lamini were the continuous fine-tuning platforms of the previous wave. Predibase was acquihired by Rubrik and Lamini by AMD, after fine-tuning took longer than expected to find adoption.
  2. A memristor is a circuit element whose resistance changes with the current that has passed through it, and it keeps that value even when the power is off. Because the value lives in the device, a grid of memristors can store weights and compute with them in the same spot.
  3. Predibase founders Piero Molino, Travis Addair, and Devvret Rishi, and Lamini founders Sharon Zhou and Greg Diamos, had experience scaling frameworks such as Horovod and Ludwig and doing hardware-software co-design. That background let them solve data routing and memory problems that block a product at scale, beyond a research proof of concept.