This episode of Unsupervised Learning features Jacob Efron interviewing Jo, CEO of Typescape (makers of Jev), about a month after Jev’s viral launch. They discuss Jev’s research philosophy centered on reliability over benchmarks, the limitations of current coding agents (especially the “tyranny of the KV cache”), future model shapes including multimodality, competition dynamics, Anthropic’s upcoming IPO, AI safety, and realistic productivity gains from coding agents.
Jev’s founding philosophy and the RLCD research direction
Jo describes himself as not mission-driven in the traditional OpenAI sense, but driven by bridging the gap between AI’s raw intelligence and its actual utility in productivity statistics.
The core insight: LLMs are API modules called by software harnesses, and the vast majority of calls will be for machine consumption (bits, structured data) not human consumption (prose, style).
Jev works backward from this “machine-native” future, building models optimized for reliable structured output rather than human-pleasing text.
Early confidence came from believing AI is clearly smart and the future is machine-native, even without early benchmark signals.
Jo emphasizes they launched not for timing but because the model was “impressively smart at a lot of things if you prompt it right” and they wanted real-world feedback to co-evolve use cases.
System one models, jaggedness, and reliability as north star
Jev models are explicitly “system one models” — fast, intuitive, pattern-matching, and prone to specific failure modes, which Jo argues is the honest characterization of how compressed intelligence actually works.
Most real-world needs are system one tasks (e.g., accounting automation) not millennium prize problems; the industry’s focus on extreme benchmarks distracts from this.
Jaggedness (unpredictable capability gaps) is not just a roadmap item but a fundamental property; Jo wants developers to complain more about specific intelligence failures so they can debug them.
Current developer behavior — building mental models of which model to use for which task — is “sad” and disempowering; the north star is SQL-like reliability where the provider becomes interchangeable.
Jo hates the dependency-update treadmill; models should work reliably in production for months without needing upgrades.
The tyranny of the KV cache in coding agents
Almost all coding agent UX decisions are constrained by the KV cache: append-only state, no true routing, sub-agents forced to pass massive context, global-variable-like state management.
This prevents advanced patterns: multiple agents sharing read/write context, live visualization dashboards separate from the main agent, semantic compaction that preserves task-relevant history, cross-session knowledge sharing across team members.
Models are trained for the KV cache append-only pattern, so deviating causes looping and degradation; Jev’s original architecture resembled LangChain but failed because base LLMs aren’t optimized for discrete routing decisions.
The goal: intelligence as composable primitives (Unix philosophy) not monolithic “write-only” frameworks.
Multimodality and future model shapes
Multimodality inherently degrades text reliability; Jev prioritizes automating easy text tasks before hard image/video tasks to stay calibrated.
Jev’s image model is “maybe 10% of the way” compared to text; Jo would be embarrassed to call it production-ready.
Exciting image/video directions: video game automation (Jo’s background), reliable computer use as interface for non-experts, “slop stoppers” to filter AI-generated noise.
New model shapes will emerge (Jev plans more), but Jo warns the field is over-focused on “flag-planting” new shapes rather than utility/reliability; this burns developer goodwill.
Competitors releasing benchmark-optimized models “dig our moat for us” by burning developer trust in benchmarks; Jev’s moat is reliability facts.
Jevans paradox (rebound effect): Jev sees demand increase as models get cheaper/better, but not yet backfire (total budget increase) — that requires “so good and so cheap I’m throwing it everywhere.”
Raised capital not for runway (service runs profitably) but to get the best AI people on the board; explicitly avoids subsidizing to create lock-in.
Benchmarks are “so bad for developers” — they optimize for demo-able, non-reliable tasks; the field has developed “extreme fatigue” with benchmarks faster than Jo expected.
Anthropic IPO and coding agent value
Coding agents are massively valuable but unclear how much use is truly net-positive vs. FOMO-driven; the market could go “very up or very down.”
Jo estimates 20-40% average speedup for engineers, with high variance: infinite% for previously impossible tasks, near-zero for ill-suited tasks.
Code is unique: git provides safeguards, revertability, human-in-the-loop; most knowledge work lacks equivalent side-effect protection (e.g., drive-thru automation still unsolved).
Anthropic’s greatest strength may be financial engineering (deal structuring on compute), not models/product/distribution; expectations may be miscalibrated by “building God is infinite” narratives — “let them cure cancer” risks becoming “let them eat cake.”
AI safety: ASI vs. misuse vs. reality
AI safety is not a monolith: ASI/RSI (recursive self-improvement) concerns are “divorced from reality of technology” — Jo estimates two gigantic breakthroughs away from true agency.
Paperclip maximizer scenarios differ fundamentally from “bad actor with super tools” threat models; the latter is more tractable and distinct.
RSI arguments often ignore Amdahl’s law and the multidimensionality of intelligence; a 10% AI research speedup that doesn’t compound isn’t RSI.
Jo is open to evidence but views ASI risk as more philosophical than technological; treating sci-fi (e.g., Superintelligence) as reality is worrying.
Call to action for builders
The field needs to “untrain” the passivity LLMs taught — developers should stop being “victims” of model providers and start building with agency.
Jev wants to enable “playable AI”: rethinking software, interfaces, primitives; real builders collaborating with model makers to unlock sci-fi use cases.
The opportunity is “incredible” if the ecosystem shifts from consumption to co-creation.
