By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

The AI Frontier: from open weights to open research — Eiso Kant, Poolside AI

From spending $12 million building language models for code before the world cared to creating a Model Factory that can take a model from pre-training to release in eight weeks, Eiso Kant has spent more than a decade betting that code is…

Eiso KantguestswyxhostVibhuhost

Jul 22, 20261h 56mWatch on YouTube ↗

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Why tool calls and MCP feel like a dead end for long-horizon agents
The conversation opens with a strong claim from Eiso Kant that putting tool-call layers between a model and its working environment is the wrong abstraction for complex tasks. He describes a shift toward giving models a virtual machine and letting them write and run code directly to accomplish goals.
- •Eiso Kant argues that MCP and tool-call layers are unnecessary overhead for complex tasks
- •He says models increasingly work better when given a VM-like environment with binaries, a codebase, and a place to write memory
- •He describes a trend away from stuffing many tools into prompts and toward code-driven interaction
- •He frames code execution (conditionals, loops, scripts) as more general than chaining tool calls
From Karpathy’s RNN essay to building early code models pre-Transformers
Eiso Kant recounts how reading Andrej Karpathy’s article on recurrent neural networks led him to pivot his startup toward neural approaches for code. He describes the pre-Transformer era as a time of experimentation—often without clear validation from the market.
- •Eiso Kant credits Andrej Karpathy’s RNN article as the trigger for his pivot into the space
- •He pursued RNNs/LSTMs and later Transformers with the belief neural nets could generalize broadly, including to code
- •He built Sourced as an open-source effort focused on machine learning and language models for code
- •He describes the period as one where “no one cared,” despite years of work
The $12M lesson: failing to scale, then feeling vindicated by ChatGPT
Eiso explains that his earlier company ultimately failed after spending significant time and money, largely because they didn’t realize scaling was the key. When ChatGPT arrived later, it felt like vindication that they had been on the right path—even if too early and too small.
- •He says the team missed the critical insight to “just keep scaling”
- •He describes the failure as the biggest of his career and mentions burning $12M with a 40-person team
- •He stepped away from language models for roughly two years, then ChatGPT changed perceptions
- •He describes renewed interest from others who resurfaced old decks and talks after ChatGPT’s release
Choosing openness again: preferring 100 foundation-model companies over five
Eiso describes Poolside’s early focus on capabilities and reinforcement learning rather than open source, then a later return to the openness question. He explains the motivation as avoiding a future where only a handful of companies control “all of intelligence.”
- •Eiso says Poolside initially didn’t prioritize open source and focused on execution
- •He describes two early premises: capabilities will keep compounding and RL will drive major capability gains
- •He explains a “fork in the road” concern: a small set of companies controlling intelligence feels dystopian to him
- •He states he’d rather live in a world with 100 foundation-model companies than five—even if Poolside were among the five
- •He lists unresolved questions: open-source business models, misuse risk thresholds, and government response
Open weights vs open research, plus Poolside’s global-remote team origin story
The group distinguishes between releasing weights and sharing research that helps others reproduce or learn. Eiso clarifies Poolside has always been an American company, but intentionally hired globally and built a remote culture to win a talent war and build resilience.
- •Eiso says weights alone don’t let others recreate what you’re doing; research-sharing can be more meaningful
- •He corrects the assumption that Poolside is a European company, saying it has been American from day zero
- •He describes a deliberate decision to avoid Bay Area-only hiring and recruit globally (US regions and internationally)
- •He recounts building training infrastructure from scratch and early painful debugging (including weeks spent on an optimizer issue)
- •He says the remote, engineering-heavy culture slowed them early but now increases release cadence
The “model factory” thesis: model building is 90% engineering
Eiso lays out Poolside’s core operating belief that successful model building depends primarily on engineering systems and process. He describes an industrial pipeline from data to distributed training to post-training and RL, and argues the real product is the repeatable factory, not any single model.
- •Eiso asserts model building is “90% engineering,” with researchers spending most time writing code
- •He describes model building as an end-to-end process: ingest, clean/transform, train, post-train, and RL
- •He introduces the “model factory” goal: shorten the cycle from researcher idea to trusted experimental result
- •He compares building the factory to building a manufacturing system like Foxconn or SpaceX’s factory
- •He gives cadence examples: five to eight weeks from pre-training start to launch for recent models, then immediately starting the next run
Agents inside the factory: automated experiments and early hints of RSI
Eiso describes agents increasingly running parts of Poolside’s research loop: writing code, launching jobs, and evaluating results. He frames this as an early glimpse of what recursive self-improvement could look like, with humans still driving ideas and debugging.
- •Eiso says agents are now a default presence on researchers’ screens, assisting with coding and job orchestration
- •He describes agents evaluating results and iterating within the research loop
- •He says the most profound impact is currently on data pipelines (pre, post, and synthetic data)
- •He notes agents are starting to influence architecture-side work as well
- •He characterizes this as “twinklings” of RSI while humans remain in the driver’s seat
Streaming data + immutability: reproducibility as the foundation of scientific rigor
Eiso explains why Poolside moved away from packaging datasets for clusters and toward streaming data into training. He connects this to an immutable data layer and versioned code that enables full traceability and reproducibility, which then supports more rigorous ablation-driven science.
- •He criticizes the common approach of packaging/tokenizing and copying datasets to clusters before training as slow and brittle
- •He describes streaming data into training so runs can begin while data is still materializing
- •He mentions a mixing service (“Blender”) to configure dataset proportions, repetition, and shuffling per run
- •He explains treating the data layer as immutable and everything as versioned to enable full traceability down to token-level cursor positions
- •He says the team evolved from “YOLO runs” to more rigorous science once reproducibility was guaranteed
Encouraging more foundation-model companies—and demystifying what’s “hard”
Eiso argues the field needs more competitors and that many people overestimate how mystical model-building is. He frames most progress as improving data or compute efficiency and says the work is often disciplined engineering and experimentation rather than theorem-proving.
- •Eiso explicitly calls on researchers to leave and start competing foundation-model companies
- •He says model building can be summarized as improving data or improving compute efficiency (with an oversimplification caveat)
- •He argues foundation-model work is mostly solid engineering: code, experiments, plots, and iteration
- •He gives an example of an engineer on Poolside’s team becoming an RL researcher within months
- •He warns there’s a limited window to help many labs catch up before RSI makes catching up harder
Laguna S and the case that persistence can matter more than raw intelligence
Eiso describes internal reactions to Laguna S, emphasizing behavior changes like verification and backtracking. He suggests that for much of knowledge work, persistence and robustness may matter more than cross-domain genius, and that this could shift where the “optimal” model size sits for many tasks.
- •Eiso quotes Peng Meng describing gains as coming from behavior: more verification, less premature victory, more persistence
- •He says Laguna S surprised them in capability and became the first release meaningfully contributing to their own work
- •He highlights sustained personal usage (reading trajectories and traces) to understand model behavior
- •He proposes that many knowledge-work tasks benefit more from persistence than extreme intelligence
- •He suggests this could support commoditization and strengthen the case for open models, while still insisting scaling remains necessary
Reinventing pre-training: RL moving earlier, curriculum design, and limits of “more environments”
The discussion turns to whether next-token prediction is enough and how reinforcement learning might move earlier into training. Eiso argues “mid-training” is mostly curriculum design and says organizational structure can artificially create discrete stages; he expects more continuous curricula over time.
- •Eiso rejects the idea that pre-training is “done” and argues there’s still major room to improve learning from the web
- •He predicts RL will move earlier than post-training—potentially earlier than what’s currently called mid-training
- •He says the web contains huge variance in quality and that the industry is not extracting enough value from it
- •He questions whether the path to AGI is simply “more environments,” suggesting more generalization from the web may be possible
- •He reframes mid-training as a crude, two-stage curriculum and predicts curricula will become more continuous as compute and experimentation improve
Low-precision training, compute bottlenecks, and why RL wall-clock time is the limiter
Eiso discusses excitement around low-precision approaches (including ternary and NVFP4-related ideas) and the engineering trade-offs they imply. He also argues that headline training-run costs miss the real expense—R&D, infrastructure, and iteration—then points out RL as the current wall-clock bottleneck due to batch-size constraints.
- •He says many architectures can work; the key question is compute efficiency
- •He expresses interest in extremely low-precision training and notes Laguna S was trained in FP8 (with exceptions like all-to-all choices in a newer run)
- •He argues “the training run is not the expensive part”; the preparation and systems are
- •He says scaling progress is tied to hardware and networking generations (training on newer systems enables larger models)
- •He identifies RL as his biggest wall-clock bottleneck because you can’t scale it arbitrarily due to batch-size/task constraints
- •He floats the idea of disaggregating compute (prefill/decode style thinking) to make RL more efficient over time
Harnesses, agents, and the path to AGI: why Poolside optimizes for minimalism
They discuss the interplay between model capability and the agent harness used to deploy it. Eiso explains Poolside’s focus on coding and long-horizon tasks as an AGI path, describes multi-harness “polishing,” and argues that minimal harnesses and general capability beat tool-heavy prompts over time.
- •Eiso frames Poolside’s focus as racing toward AGI, with coding as a forcing function for long-horizon reasoning and tool use
- •He says models can feel qualitatively different within minutes of a new checkpoint, beyond benchmark results
- •He describes “multi-harness polishing” to transfer behaviors across harness formats, while keeping most RL focused on their own harness
- •He argues adding many external harnesses into RL increases complexity and reduces scientific clarity about what improved the model
- •He advocates for a minimal harness (few tools) and expects tool-stuffed system prompts to fade as models write code directly in environments
Vision and multimodality priorities: why language still dominates their focus
Eiso says Poolside does not yet have vision understanding in these models but considers it the next key step, while de-prioritizing audio due to focus constraints. He argues language is uniquely dense for encoding knowledge and reasoning, making it a compute-efficient place to push capabilities first.
- •Eiso says vision understanding is important and Poolside has started efforts, but it’s not in these models yet
- •He says audio is not a priority for Poolside, not because it’s hard but because focus matters
- •He argues language is an extremely dense medium for knowledge and reasoning compared to images/video in terms of signal-to-noise and compute efficiency
- •He praises maniacal focus on scaling capabilities, while acknowledging modality-focused labs are doing exciting work too
- •He expects different foundation-model companies to pursue different modality priorities
Scaling roadmap, naming stories, and raising $500M
The conversation covers Poolside’s model lineup and ongoing scaling (including a new, larger medium model training run). Eiso also explains how the name “Poolside” came from a moment at a conference and why it symbolizes choosing harder paths, then discusses how investor belief in AGI and competitive dynamics shaped their fundraising environment.
- •Eiso says a new, larger Laguna medium started training and mentions plans for larger sizes (L/XL)
- •He emphasizes that small/medium naming is relative and jokes that naming is the hardest problem
- •He tells the origin story: “Poolside” came from being told to move a sensitive conversation to a restaurant poolside during a conference
- •He says the name reminds them not to round down ambitions and to choose the harder path
- •He says fundraising changed as more investors accepted that model capabilities will keep compounding and that product-market fit is proven
- •He frames the investor divide as belief in an intelligence oligopoly vs belief that more winners can emerge
Open models, AI safety, democracy, and avoiding an intelligence oligopoly
Eiso argues that unilateral decisions by a few companies concentrate power and prefers democratic governance as capabilities rise. He supports openness today while acknowledging that at higher capability levels some restrictions may be necessary, and warns against blanket regulations that could entrench incumbents and reduce competition.
- •Eiso says concentrated unilateral control by a few companies is dangerous and prefers democratic processes to shape policy
- •He distinguishes misuse-focused safety from speculative doomsday scenarios and supports better evals and safety institutes
- •He says current model capability levels don’t justify restricting open models and warns it would slow innovation
- •He explicitly rejects extreme positions: “open forever” vs “close everything now,” arguing nuance is required
- •He warns that broad restrictions can create oligopolies, using the example of cigarette advertising bans entrenching incumbents
- •He objects to model-provider restrictions that prevent using models for foundation-model development, calling the situation “wild”
Engineering productivity, hiring for agency, and Poolside’s recruiting pitch
In closing, Eiso connects productivity to cycle time: converting user value into shipped software quickly with fewer people involved. He describes “agency” as the key trait to hire for, argues constraints create innovation, and invites candidates across applied research and engineering roles due to the potential for high individual impact.
- •Eiso frames engineering productivity as reducing lead time from valuable idea to shipped outcome, with fewer people in the loop
- •He says AI tools compress iteration cycles dramatically, enabling smaller teams and more direct builder-to-customer workflows
- •He identifies agency as the core interview/hiring signal and says it shows up in past behavior
- •He argues high-agency teams require clear shared goals and constraints (“lanes”) to align exploration
- •He says Poolside is hiring broadly across applied research and engineering, emphasizing high individual impact in a small org

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.