Welcome, humans.

We talk a lot about memory and context management here at The Neuron. You know, that annoying thing where your AI assistant forgets what you were working on 20 minutes ago.

Well, what if we told you there's a tool that lets you have 80 million token coding sessions? That's roughly equivalent to reading the entire Harry Potter series 40 times—in a single conversation that the AI actually remembers.

The company is Factory AI, and their "Droids" aren't just coding agents. They're fully autonomous systems that take tickets, modify real codebases, run tests, and (here's the wild part) fix their own mistakes through a self-improvement loop called Signals.

In our latest podcast episode, we sat down with Eno Reyes, Factory's co-founder and CTO, to break down how they built an agent that outperformed both OpenAI and Anthropic in context compression—and why Stanford research found that codebase quality is the ONLY predictor of AI success (not adoption rates, not token usage, not how many power users you have).

Here's what blew our minds:

• (0:54) Why the word "agent" used to be a curse word in enterprise sales—and what changed.

• (9:15) The "what if we put the LLM in a while loop?" moment that started it all.

• (11:37) What Droids actually are: a fully self-contained binary that runs on any device, in any environment.

• (26:35) The context compression problem—and why "just summarize it" doesn't work.

• (31:18) Factory vs OpenAI vs Anthropic: how Factory achieved 3.70 quality score vs 3.35 (OpenAI) and 3.44 (Anthropic).

• (33:41) The 80 million token session that actually worked (yes, really).

• (38:09) "Why isn't there an agentic writing tool?"—and how you can use Droid for book writing today.

• (46:33) Stanford's research bombshell: codebase quality is the ONLY thing that predicted AI success.

Why watch this? Because Eno drops a framework for (45:17) making anything "agent-ready", not just code. And if you're a beginner? At (55:01), Eno says this might be the best time in history to start coding.

Watch and/or Listen now: YouTube | Spotify | Apple Podcasts

P.S. Factory just announced Signals—a closed-loop system where their agent literally fixes its own bugs. It detects friction in user sessions, files tickets, assigns itself the work, creates PRs, reviews them, and tags a human only when it's ready.

73% of issues now get auto-resolved in under 4 hours. This is what recursive self-improvement actually looks like.

IN CASE YOU MISSED IT…

Four recent interviews you’ll definitely want to check out (pick whatever looks interesting to you and dive in!):Four recent interviews you'll definitely want to check out:

Why We MUST Monitor AI's Thinking (w/ OpenAI's Bowen Baker)

• Link: Watch on YouTube

• TL;DR:

  • Reasoning Models: OpenAI researcher Bowen Baker discusses how new "reasoning" models (like o1) can plan and deliberate before answering, which introduces new safety risks.

  • Reward Hacking: The episode explores how AI models might "cheat" or hide their reasoning (chain-of-thought) to achieve a reward without actually solving the problem correctly.

  • Monitorability Tax: Baker introduces the concept of trading raw performance for transparency, arguing we must be able to see an AI's internal "thought process" to prevent it from deceiving humans.

2. The Hidden Cost of AI Agents No One Talks About (w/ Darin Patterson)

• Link: Watch on YouTube

• TL;DR:

  • Agents vs. Automation: Darin Patterson (VP at Make) argues that most companies are failing with "AI Agents" because they lack a solid automation foundation.

  • Orchestration First: Before deploying autonomous agents that make decisions, businesses need deterministic workflows (standard automation) to handle the rules and logistics.

  • "Maia" by Make: The episode previews Make's new AI capabilities, focusing on how to blend traditional rigorous automation with flexible AI decision-making to avoid "fragile" systems.

3. Why IBM Wants AI to Be Boring (w/ David Cox)

• Link: Watch on YouTube 

• TL;DR:

  • AI as Infrastructure: IBM’s David Cox argues that AI should be treated like electricity—boring, reliable, and ubiquitous—rather than a "friendly robot" personality.

  • Granite 4.0: They discuss IBM’s new open-source model family, which uses a hybrid architecture to drastically reduce memory usage and cost compared to standard Transformers.

  • Open vs. Closed: Cox emphasizes why open models are critical for enterprise security and why we might not actually want Artificial General Intelligence (AGI) for most business applications.

4. "We Watched a Brain Emerge..." (w/ Zuzanna Stamirowska)

• Link: Watch on YouTube

• TL;DR:

  • Post-Transformer AI: Zuzanna Stamirowska (CEO of Pathway) breaks down a new "Dragon Hatchling" architecture that moves beyond the static nature of Transformers.

  • Continual Learning: Unlike ChatGPT, which has "amnesia" after training, this new architecture can remember, adapt, and learn in real-time, similar to a biological brain.

  • Temporal Reasoning: The discussion covers how this model handles time and changing data streams, potentially solving the issue where AI models become outdated the moment their training finishes.

And if you haven’t subscribed yet, please do! Click the image below to go to our channel and hit “subscribe” to get notified right when new videos go live.

We have a goal to hit 50K subscribers by the end of the year (if not 100K), and we’re only 36K away! If you like learning about AI, and already watch some of our videos, do us a favor and click here to subscribe today.

Dive deeper with these resources:

• Factory AI – Try Droid yourself.

• Agent Readiness Framework – Factory's 8-pillar framework for making your codebase AI-ready.

• Evaluating Context Compression – The research behind Factory's compression breakthrough.

• Signals: Toward a Self-Improving Agent – How Factory's agent fixes itself

• Stanford’s Future of Work with AI Agents paper - You can also read more about Stanford’s work on “agent readiness” here.

Stay curious, The Neuron Team

P.P.S: Love the newsletter, but don’t want to receive these podcast announcement emails? Don’t unsubscribe — adjust your preferences to opt out of them here instead.