Personal AI and Continual Learning: New Frontiers in Agentic AI

The slides failed as he took the stage. He offered to improvise the whole talk and kept going for about a minute until they came back. His closing line: "By the way, this whole thing was written by my personal agent, because I was pretty lazy today."

TL;DR

• Coding agents succeeded because rewards are verifiable: the agent writes code, a unit test judges it, the reward signal feeds RL, and the whole reinforcement-learning toolbox makes the agent extremely powerful. The catch: the next domain has no unit test.
• Three directions he sees: scientific discovery and self-improvement, automating the economy, and — his own passion — personal AI: a digital entity that accompanies you, understands you deeply, and reaches out proactively. Possibly the next computing paradigm after the PC, the phone, and the internet.
• Five unsolved problems: (1) running RL for the personal-AI use case itself (today's personal agents are coding agents generalized sideways); (2) personalization; (3) memory and long context, for tasks spanning days; (4) privacy and security, e.g. prompt injection; (5) cost — an agent running 24/7 easily reaches thousands of dollars a month.
• Two philosophies of memory: in-context (discrete facts — a phone number, a past conversation) versus weight-level updates (implicit knowledge — preferences, the style someone writes emails in). Each fails where the other works, so you need both.
• Also a window to hand control back: along the mainframe → PC → smartphone arc, cloud AI is a step backwards. Personal AI could put the model weights and your private data on a device in your own home.

Key Points

What comes after coding agents? (~00:43–00:46)

He starts from the year's most visible fact: coding agents have enormous economic appeal — you can build an extremely profitable company purely on serving them. The reason is structural. Coding has verifiable rewards: the agent writes code, a unit test or equivalent decides whether it was right, that becomes a reward signal, and RL takes it from there. Everyone agrees coding is one of the fundamental building blocks you solve before moving to other domains.

So what's next? Three directions: scientific discovery and self-improvement (echoing the two speakers before him), automating the economy, and personal AI.

His picture of personal AI isn't a chatbot. It's a digital entity that accompanies each of us, understands us deeply, proactively reaches out, and makes life happier and smoother. He's willing to call it a candidate for the next computing paradigm after the personal computer, the mobile phone, and the internet.

He's candid about where it stands: early systems like OpenClaw exist, but anyone who has played with them knows there's no mass-market appeal yet — you have to tinker, and you have to maintain them carefully.

The five unsolved problems (~00:46–00:48)

1. RL for the personal-AI use case. Today's personal AI systems are frequently coding agents generalized into a use case they weren't trained for.
2. Personalization. An agent running 24/7 in the background, acting autonomously, had better understand precisely how to help you.
3. Memory and long context, because tasks can span a full day or several.
4. Privacy and security. His example lands: imagine a friend — or an enemy — texting your personal agent a string that takes it over and hands them all your secrets.
5. Cost. Add up the tokens an always-on background agent consumes and you reach thousands of dollars a month, which is prohibitive for most people.

The ownership opportunity (~00:48–00:49)

He places this in the arc of computing: mainframes → personal computers → smartphones moved control toward the individual, and cloud AI is a step back the other way — you don't own the physical device and you don't own the means of inference, so control sits with the companies serving LLM APIs.

Personal AI is a chance to return to that regime: a physical device in your home or office holding the model weights and storing your private data — more secure, better for privacy. He's careful about the claim: it isn't guaranteed, but it's an opportunity to build something genuinely for the benefit of the individual.

Direction 1: two philosophies of agent memory (~00:49–00:50)

• In-context memory: a memory system the agent queries, with information moving as tokens or text. Good for retrieving a phone number or a past conversation.
• Weight-level updates: changing the model itself.

Their weaknesses are complementary. At the weight level it's genuinely hard to store an individual fact like a phone number or someone's name. In the in-context world it's hard to store implicit knowledge — what my user's preferences are, what style they write their emails in, whether I can imitate it. That's very hard to specify in a text file. His conclusion: we need both, and his slides list research papers pursuing each.

Direction 2: two approaches to personalization (~00:50–00:51)

The problem is aligning the model not with the whole user population but with the individual it serves — which he considers essential if the agent is going to reliably represent you and get things done in the background.

• Passive: observe what the user does and leverage the data they generate in the course of their life.
• Active: probe the user about their preferences, or optimize directly against their feedback — I say "this was good, this wasn't," and the system updates weights or prompts accordingly.

Direction 3: infrastructure — multi-tenant LoRA and local hardware (~00:51–00:53)

Personal AI may need a different toolset and newly deployed infrastructure. Two directions are useful today:

Multi-tenant LoRA. Training used to mean updating all the weights, which now number in the trillions. LoRA adapters represent a small fraction of the weights and are the only thing modified during training, which makes post-training and RL far more efficient — and opens up one custom fine-tuned model per user, plus continual-learning experiments. River AI has built and released its own multi-tenant LoRA system.

Local hardware. Moving inference from the data center to a device in your hand, home, or office. People are only starting to explore this, and there's a lot of low-hanging fruit in optimizing local inference; he thinks it might one day rival data-center inference. The bottleneck is model capacity: fitting trillions of parameters into a small form factor and running inference efficiently without overheating your house. Today that takes a whole rack of GPUs; he's hoping for a single device.

He closes with a plug for river.ai/api — and the line about his personal agent having written the talk.

Quotes

"Imagine that your friend — or your enemy — texts your personal AI agent a string that then allows them to take it over, and they can get all your secrets." (~00:47)

Prompt injection, rendered concrete in the personal-AI setting.

"Cloud AI is sort of a step away from giving the individual control. You don't have the physical device. You don't have the means for doing inference." (~00:48)

Why he treats personal AI as an ownership question as much as a capability one.

"By the way, this whole thing was written by my personal agent, because I was pretty lazy today." (~00:53)

The last line of the talk.