Cars24 Recovers 12% of Dead Leads as Wary Developers Keep Feeding LLMs More

01A language with no market is training its own speech models, the same week open models chased closed systems on video

In New Zealand's far north, a Māori broadcaster is training speech models for te reo, a language too small for any vendor to serve, according to a report circulating on Hacker News. The data stays under a license that keeps it with the community. No cloud provider offered to build this. The broadcaster owns the result instead of renting it.

The same report lists others running that play. PwC fine-tuned an open model on financial language and now serves hundreds of clients from its own hardware, with no per-token meter. A Lausanne team built an open medical model with the Red Cross, tuned to humanitarian guidelines, with clinical trials planned at home and in Tanzania. In East Africa, farmers diagnose cassava disease on a model that runs offline, on the phone itself. A Swiss public consortium trained a national model on public supercomputers, then released the weights, data, and training code.

What links them is not scale. Each serves a use case the commercial market declined to build, and each keeps ownership rather than paying a vendor to rent it back.

That is one front. The same period saw open source pushing on a second one: the capability gap with closed labs. Two papers posted to Hugging Face make the point.

VideoChat3 ships as a fully open video model, including the training components most open releases withhold. Its authors argue existing open video models generalize poorly across video types and demand heavy compute, and position VideoChat3 as an efficient generalist against that limitation.

Boogu-Image-0.1 states the comparison outright. Its authors benchmark the model family against closed systems like Nano-Banana-Pro and GPT-Image-2, then note those reach their results through system-level integration rather than a single model. Boogu claims competitive text-to-image generation, instruction-based editing, and bilingual Chinese-English rendering across Base, Turbo, and Edit variants.

The two fronts move at once. Markets abandon the long tail, and open projects pick it up. Closed labs hold the frontier, and open releases keep narrowing what those labs can charge to gate.

02Cars24 recovered 12% of dead leads by grading its AI on cost per successful task

When a used-car shopper stops replying halfway through a chat, Cars24's system does not file the lead away as lost. A voice or chat agent picks the conversation back up. The company says those agents now handle more than one million conversation minutes a month, and that the follow-ups claw back 12% of leads it would otherwise write off.

Cars24 has pushed the same agentic workflows to teams across the company, according to an account of the deployment published by OpenAI. That is the easy part to announce. The harder part, for any buyer, is proving the spend returns more than it costs, and the platform's answer is to count outcomes rather than model capability.

That framing lines up with a measurement pitch from Sarah Friar, OpenAI's CFO, who has laid out an "AI scorecard" built on four things: useful work completed, cost per successful task, dependability, and return on compute. The metrics are the vendor's, but they describe a buyer's problem. A model that answers well in a demo is not the same as one that closes a sale cheaply and reliably at a million minutes of load.

Cost per successful task is the line that matters most to a deploying team. It ties a price to a result a business already tracks, in Cars24's case a recovered lead, instead of to tokens processed or a benchmark score. Dependability sets a second bar. An agent that recovers leads 95% of the time but hangs up on the rest carries a support cost that eats the gain.

The center of gravity in these decisions has moved. The question is no longer whether a model can hold a natural voice conversation, since it plainly can. It is whether routing a million monthly minutes through one pays for the compute, the integration, and the failures. Cars24's 12% recovery figure is the kind of number a finance team can put against an invoice. The next thing worth watching is whether that rate holds as the workflows spread to teams whose tasks are harder to score than a missed sales lead.

03The developers who trust LLMs least keep handing them more

Software engineers who publicly concede the case against large language models are, in the same breath, granting those models wider autonomy. Two items surfaced this week hold both halves of that split.

A developer's blog post, "The LLM Critics Are Right. I Use LLMs Anyway," opens by conceding nearly every objection to the technology, then reports using it heavily regardless. The author describes attending Local-First Conf in Berlin, where the dissonance was on stage. Armin Ronacher, who created Flask and was an early Sentry engineer, now runs Earendil and builds Pi.dev, an open-source coding agent harness. Asked live how the project handles a flood of LLM-generated pull requests, Ronacher said they auto-close almost all PRs and issues, according to the post. He added that people shouldn't be discouraged from opening them, because "the human will always shine through." The company's own purpose page reads: "In a world hurtling towards AI, we believe humans are the best agents." From the audience, the author counted a room full of laptops with Claude Code open while speakers voiced their doubts.

The second item drops the doubt entirely. A team at tryai.dev built an agentic harness that hands a model a song, a fixed dollar budget, and a set of tools, then steps back. The model researches which video generators exist, picks ones on FAL or Replicate, generates clips, watches its own footage, and edits with ffmpeg through a local shell. Its only thinking tool costs nothing; only image and video generation spend budget. The team ran Claude Fable 5 and GPT-5.6 Sol at $25 and $100 each, four runs total, every run fed the same track, Bruno Mars and Mark Ronson's "Uptown Funk." Every tool call was logged; the two models diverged in what they chose to research, generate, and cut.

Nobody in either item claims the models are reliable. The blog author still ships with them daily. The harness builders still let a model spend real money unsupervised across a long-horizon task. The gap between stated skepticism and delegated authority is not closing on the critic's side. It is widening on the practitioner's.