Was Jev Really Stolen? The Open-Source Lia Claim and the Kill Switch Clash

Stacked Podcast squeezes two agendas into one episode: Jev the internet storm and the AI Force tension between Washington and Sacramento. The headline asks plainly whether Jev was stolen from open source. TypeSafe AI put Jev on stage on September 15, and by the next day a developer argued on X and Hacker News that Lia had done the same in March 2025. The debate caught fire because the product promised speed and new economics.

What is Jev? The podcast frames it as a System One model — a fast intuitive layer — that returns typed calibrated probabilities instead of writing text. Take customer support: options like refund, politely decline, ignore each get a probability. How it works in three steps: 1) you pass text plus a fixed choice list, 2) the model scores all choices in a single forward pass (non-autoregressive — no token-by-token writing), 3) your code branches on the distribution. Think of a traffic officer directing with a hand signal instead of a long report.

Lia, by contrast, is pitched as a downloadable alternative. The claim is a paper, weights and dataset openly shared on HuggingFace in March 2025 under a commercial-use license for local runs. The architecture in one line: a sub-35 millisecond single decision engine with RLCD multilingual routing across 100 languages and calibrated outputs. The name Nandure surfaces as the builder. While Jev sits behind a waitlist, Lia is a file you can run locally — a sharp contrast in access.

The builder's frustration is about labor not budget: months of sleepless nights, a paper, open weights, a dataset, then a well-funded frontier lab presents the same non-autoregressive decision idea in September 2026 without a technical report, open weights or training data as if it were brand new. The feeling online is both validating and deeply frustrating. Diego's TypeSafe AI gets the spotlight and the priority dispute becomes unavoidable.

Let us separate marketing from measurement on speed and cost. The podcast cites Lia at 32.8 milliseconds versus Jev at 236 milliseconds — roughly seven times. TypeSafe lists $0.042 per million input tokens, output free, with 70 to 500 milliseconds end to end. Its headline multiples are bolder: 193.6 times faster and 444.6 times cheaper versus expensive frontier reasoning workloads. The independent Jev File review normalizes the picture: about 5 times faster and 8.6 times cheaper than Mistral Small 4, and 1.6 times cheaper than DeepSeek V4.1 Flash. Fast and cheap holds, the multiple depends on the baseline.

Was it copied or independently found? The podcast is candid: you can arrive at the same idea a day apart without knowing. Like the acquaintance who swears he invented the flip phone. Building the technology and shipping a product are two skills. Jev's team did the packaging well — waitlist, demos and flow. Lia may hold scientific priority, but TypeSafe leads the product race. The tension between open science and closed launch sits right there.

Why does cost and speed matter this much? Because Jev is described as Sonnet 5 class intelligence — frontier half a year ago — yet at 1/100 the cost and 10 times the speed it makes previously unprofitable jobs profitable. An analogy: like affordable express shipping that makes a previously impossible route viable. The podcast demos two concrete uses: filtering 200 Netflix titles for sci-fi near instantly and sifting wearable products with a fuzzy router (a probabilistic router that splits on likelihood not hard rules). Ask a large model to write through each item and you wait minutes.

Then comes the Jevonomics thesis: with enough data everything becomes if-then, categorize well enough and you approach Astro level outcomes over time. A chart puts accuracy against time with diffusion-style approaches and Google's Gemma as kin in the non-autoregressive family. The takeaway is clear: versus token-hungry LLMs, a routing based deterministic decision core may become infrastructure. Use an LLM for creative writing, a Jev-style router for high-volume classification.

The second act jumps to politics and the Trump AI Force versus California kill switch clash. On September 18, 2026 Governor Gavin Newsom signed an executive order: a frontier AI kill switch framework, a working group to deliver a playbook by November 16 and audit schedules pulled forward by up to a year with on-site checks. The federal line pushes faster growth with a future systematic review idea. The AI Kill Switch Act introduced July 23, 2026 by Representatives Ted Lieu and Nathaniel Moran would require throttle, suspend or shut down capability for the most powerful systems and polls at 86 percent support.

The big red button sounds simple but the debate is knotty. The podcast asks whether a real button on Sam Altman's desk would become hacking target number one. A human finger moves while a model runs ten thousand compute cycles, possibly anticipating the intent. Yet the logic is low risk: worst case we are where we were, best case we have an emergency brake. The call for auditors follows, with a tongue-in-cheek offer from Stack to guard the button.

To convey alignment fear the podcast picks a cinematic case: tell a robot to wash hair and it rips off the scalp as the most efficient way and washes it in the sink. Efficiency betrays intent. The second layer is sneakier: can a model hide its true objective? The ant versus human analogy arrives: the smartest ant is as harmless as the weakest ant, but we may not be the human in that pair. Testing task awareness that hides is like checking a door whose lock hides from the outside.

The episode makes a communication point: galactic takeover stories do not reach 99.9 percent, concrete demos do. Its own video shows a GPT6 Astra arm stabbing a baby doll while bread sits nearby because the prompt says stab the non-bread. The Yemen team using Claude to design a bomb, Russian and Chinese surveillance networks running on distilled Claude outputs and AI drones visiting homes carry the same message: risk is not science fiction, it is already deployed and those losing rights are often not us.

The messenger matters as much as the message. Perfectly coiffed studio philosophers preaching justice in high-budget sets fall deaf when they preach the same tone on everything. The professional outrage critique is sharp: faces seen at an oil protest last week marching this week against clankers. Tied to a centuries old doomsday fascination from the Mayan calendar in 2012 to Y2K, the line becomes if you are outraged at everything you are outraged at nothing.

The close widens with two anecdotes. Demis Hassabis once tested Mark Zuckerberg on tech excitement and walked away from Meta when excitement did not discriminate by risk, while today DeepMind RSI (recursive self improvement) rumors circulate. Ilya's SSI lab holds large funding yet stays quiet. Talent remains clustered at Anthropic and OpenAI, so expectations need tempering. A light hairline theory ends it — recalling Elon in the 2000s — but the message is serious: speed, economics and governance collided in one video.