--- schema: 1 type: video title: 'Local Models: Trust, Control, Optimization — Carter Abdallah, NVIDIA' description: >- When Fable was pulled back and access to frontier systems stopped looking guaranteed, Lucas Atkins watched enterprises move to Chinese open models, not because they scored better but because availability could be counted on. That is his working definition of trust, and he separates it hard from safety: an open model is a directory of files you can inspect, running on code you can read, while the same claim about a closed API is unverifiable by construction. Arcee's response was to reorient the whole company and pretrain a 400 billion parameter model in six months, which he says plenty of people called impossible. The rest is control. Vincent Weisser describes a customer specializing an open model to automate finance work in a week or two and landing better results than Opus at a fraction of Haiku's cost. Closed terms of service bar you from training on outputs, so owning the model also means owning the traces, which is what makes a data flywheel possible at all; Nemotron and Trinity both adopted the open MDW license to put that permission in writing. Chris Alexiuk's framing is the mismanaged genius: a model tuned to be good across every harness is optimal for nobody's, and open weights let you fit it to the one or two things you actually do. The predictions land where you would expect from this panel, including open models reaching Fable level capability inside a year, and Atkins hoping the share of people who have ever run a model locally climbs from a rounding error to 10% to 15%. Speaker info: Carter Abdallah, moderator (NVIDIA): - https://x.com/Baxate - https://baxate.com Vincent Weisser (Prime Intellect): - https://www.linkedin.com/in/vincentweisser - https://www.primeintellect.ai/ Lucas Atkins (Arcee AI): - https://x.com/latkins - https://arcee.ai Chris Alexiuk (NVIDIA): - https://x.com/llm_wizard - https://www.alexi.uk/ Timestamps: 0:00 - Welcome and why this panel is the whole stack 1:14 - Prime Intellect: keeping the training stack open 2:15 - Arcee: why the west was losing the open model lead 3:22 - Pretraining a 400 billion parameter model in six months 4:23 - Nemotron: faster models are smarter models 6:35 - Is open source actually less trustworthy 7:40 - Trust is not safety 9:48 - When access stopped being guaranteed 10:56 - Releasing the data sets alongside the weights 11:59 - The open superintelligence stack 13:02 - Post training as the accessible layer 14:07 - Beating frontier models on a specific use case 15:11 - Making your costs predictable 16:11 - When the model and the harness blend together 18:15 - The mismanaged genius 19:19 - A call to action for builders 20:24 - Who owns the data you generate 21:27 - Owning your outputs, not just your weights 22:35 - The open MDW license 23:36 - Where post training unlocks new use cases 27:46 - You do not need frontier intelligence for most tasks 28:53 - Why efficiency has to happen in the open 29:55 - Closed models are not the enemy 33:05 - Predictions for the next year 38:22 - Running everything on your laptop 39:29 - Agent operating systems and the next Siri moment 41:36 - Closing --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/FWMJQDH3iK0-local-models-trust-control-optimization Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: 'https://liquidgold.cc/watch/FWMJQDH3iK0' markdown: 'https://liquidgold.cc/watch/FWMJQDH3iK0.md' llms: 'https://liquidgold.cc/watch/FWMJQDH3iK0/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T05:47:03.596Z' language: null children: [] videoId: FWMJQDH3iK0 durationSeconds: 2601 publishedAt: '2026-08-07T02:00:06.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 86 endSeconds: 142 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=86' heldReason: rights - ordinal: 1 text: null startSeconds: 154 endSeconds: 208 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=154' heldReason: rights - ordinal: 2 text: null startSeconds: 219 endSeconds: 273 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=219' heldReason: rights - ordinal: 3 text: null startSeconds: 283 endSeconds: 403 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=283' heldReason: rights - ordinal: 4 text: null startSeconds: 480 endSeconds: 591 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=480' heldReason: rights - ordinal: 5 text: null startSeconds: 414 endSeconds: 470 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=414' heldReason: rights - ordinal: 6 text: null startSeconds: 656 endSeconds: 667 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=656' heldReason: rights - ordinal: 7 text: null startSeconds: 783 endSeconds: 795 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=783' heldReason: rights - ordinal: 8 text: null startSeconds: 795 endSeconds: 849 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=795' heldReason: rights - ordinal: 9 text: null startSeconds: 915 endSeconds: 926 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=915' heldReason: rights - ordinal: 10 text: null startSeconds: 926 endSeconds: 982 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=926' heldReason: rights - ordinal: 11 text: null startSeconds: 992 endSeconds: 1103 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=992' heldReason: rights - ordinal: 12 text: null startSeconds: 1103 endSeconds: 1113 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=1103' heldReason: rights - ordinal: 13 text: null startSeconds: 1236 endSeconds: 1290 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=1236' heldReason: rights - ordinal: 14 text: null startSeconds: 1303 endSeconds: 1360 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=1303' heldReason: rights - ordinal: 15 text: null startSeconds: 1429 endSeconds: 1666 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=1429' heldReason: rights - ordinal: 16 text: null startSeconds: 1429 endSeconds: 1666 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=1429' heldReason: rights - ordinal: 17 text: null startSeconds: 1666 endSeconds: 1676 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=1666' heldReason: rights - ordinal: 18 text: null startSeconds: 1752 endSeconds: 1797 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=1752' heldReason: rights - ordinal: 19 text: null startSeconds: 1808 endSeconds: 1994 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=1808' heldReason: rights - ordinal: 20 text: null startSeconds: 2004 endSeconds: 2125 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=2004' heldReason: rights - ordinal: 21 text: null startSeconds: 2136 endSeconds: 2264 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=2136' heldReason: rights - ordinal: 22 text: null startSeconds: 2318 endSeconds: 2375 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=2318' heldReason: rights - ordinal: 23 text: null startSeconds: 2375 endSeconds: 2386 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=2375' heldReason: rights - ordinal: 24 text: null startSeconds: 2386 endSeconds: 2505 watch: 'https://liquidgold.cc/watch/FWMJQDH3iK0?t=2386' heldReason: rights --- # Local Models: Trust, Control, Optimization — Carter Abdallah, NVIDIA When Fable was pulled back and access to frontier systems stopped looking guaranteed, Lucas Atkins watched enterprises move to Chinese open models, not because they scored better but because availability could be counted on. That is his working definition of trust, and he separates it hard from safety: an open model is a directory of files you can inspect, running on code you can read, while the same claim about a closed API is unverifiable by construction. Arcee's response was to reorient the whole company and pretrain a 400 billion parameter model in six months, which he says plenty of people called impossible. The rest is control. Vincent Weisser describes a customer specializing an open model to automate finance work in a week or two and landing better results than Opus at a fraction of Haiku's cost. Closed terms of service bar you from training on outputs, so owning the model also means owning the traces, which is what makes a data flywheel possible at all; Nemotron and Trinity both adopted the open MDW license to put that permission in writing. Chris Alexiuk's framing is the mismanaged genius: a model tuned to be good across every harness is optimal for nobody's, and open weights let you fit it to the one or two things you actually do. The predictions land where you would expect from this panel, including open models reaching Fable level capability inside a year, and Atkins hoping the share of people who have ever run a model locally climbs from a rounding error to 10% to 15%. Speaker info: Carter Abdallah, moderator (NVIDIA): - https://x.com/Baxate - https://baxate.com Vincent Weisser (Prime Intellect): - https://www.linkedin.com/in/vincentweisser - https://www.primeintellect.ai/ Lucas Atkins (Arcee AI): - https://x.com/latkins - https://arcee.ai Chris Alexiuk (NVIDIA): - https://x.com/llm_wizard - https://www.alexi.uk/ Timestamps: 0:00 - Welcome and why this panel is the whole stack 1:14 - Prime Intellect: keeping the training stack open 2:15 - Arcee: why the west was losing the open model lead 3:22 - Pretraining a 400 billion parameter model in six months 4:23 - Nemotron: faster models are smarter models 6:35 - Is open source actually less trustworthy 7:40 - Trust is not safety 9:48 - When access stopped being guaranteed 10:56 - Releasing the data sets alongside the weights 11:59 - The open superintelligence stack 13:02 - Post training as the accessible layer 14:07 - Beating frontier models on a specific use case 15:11 - Making your costs predictable 16:11 - When the model and the harness blend together 18:15 - The mismanaged genius 19:19 - A call to action for builders 20:24 - Who owns the data you generate 21:27 - Owning your outputs, not just your weights 22:35 - The open MDW license 23:36 - Where post training unlocks new use cases 27:46 - You do not need frontier intelligence for most tasks 28:53 - Why efficiency has to happen in the open 29:55 - Closed models are not the enemy 33:05 - Predictions for the next year 38:22 - Running everything on your laptop 39:29 - Agent operating systems and the next Siri moment 41:36 - Closing --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/FWMJQDH3iK0-local-models-trust-control-optimization Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/watch/FWMJQDH3iK0 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.