Hacker News: oscarmoxonhttps://news.ycombinator.com/user?id=oscarmoxonHacker News RSShttps://hnrss.org/hnrss v2.1.1Tue, 21 Apr 2026 01:27:04 +0000<![CDATA[How we Built Private Post-Training and Inference for Frontier Models]]>Article URL: https://www.workshoplabs.ai/blog/private-post-training

Comments URL: https://news.ycombinator.com/item?id=47403488

Points: 7

# Comments: 1

]]>Mon, 16 Mar 2026 19:19:16 +0000https://www.workshoplabs.ai/blog/private-post-trainingoscarmoxonhttps://news.ycombinator.com/item?id=47403488https://news.ycombinator.com/item?id=47403488<![CDATA[New comment by oscarmoxon in "Open Weights isn't Open Training"]]>Some of this exists already in pockets (Common Crawl, The Pile, RedPajama are all volunteer/open efforts). I suppose there's no equivalent of the "edit this page and see the impact" like with have with Wikipedia. Contributing to an open dataset has no feedback loop if the training infrastructure that would consume it is closed... seems like a feedback problem.

]]>Tue, 10 Mar 2026 19:57:55 +0000https://news.ycombinator.com/item?id=47328049oscarmoxonhttps://news.ycombinator.com/item?id=47328049https://news.ycombinator.com/item?id=47328049<![CDATA[New comment by oscarmoxon in "Open Weights isn't Open Training"]]>Agree that this makes it unlikely we see frontier training data OS'd but this is a separate problem from software and infrastructure transparency, which has none of those constraints. Training stack, the parallelism decisions, documented failure modes are engineering knowledge and there's no principled reason it doesn't ship.

]]>Tue, 10 Mar 2026 19:54:32 +0000https://news.ycombinator.com/item?id=47328012oscarmoxonhttps://news.ycombinator.com/item?id=47328012https://news.ycombinator.com/item?id=47328012<![CDATA[New comment by oscarmoxon in "Open Weights isn't Open Training"]]>Agree, this feels like a distinction that needs formalising...

Passive transparency: training data, technical report that tells you what the model learned and why it behaves the way it does. Useful for auditing, AI safety, interoperability.

Active transparency: being able to actually reproduce and augment the model. For that you need the training stack, curriculum, loss weighting decisions, hyperparameter search logs, synthetic data pipeline, RLHF/RLAIF methodology, reward model architecture, what behaviours were targeted and how success was measured, unpublished evals, known failure modes. The list goes on!

]]>Tue, 10 Mar 2026 18:37:40 +0000https://news.ycombinator.com/item?id=47327148oscarmoxonhttps://news.ycombinator.com/item?id=47327148https://news.ycombinator.com/item?id=47327148<![CDATA[New comment by oscarmoxon in "Open Weights isn't Open Training"]]>Compute costs are falling fast, training is getting cheaper. GPT-2 costs pocket change to train, and now it costs pocket train to tune >1T parameter models. If it was transparent what costs went into the weights, they could be commodified and stripped of bloat. Instead the hidden cost is building the infrastructure that was never tested at scale by anyone other than the original developers who shipped no documentation of where it fails. Unlike compute, this hidden cost doesn't commodify on its own.

]]>Tue, 10 Mar 2026 18:31:28 +0000https://news.ycombinator.com/item?id=47327073oscarmoxonhttps://news.ycombinator.com/item?id=47327073https://news.ycombinator.com/item?id=47327073<![CDATA[New comment by oscarmoxon in "Open Weights isn't Open Training"]]>The framing here is undersold in the broader discourse: "open weights" is a ruse for reproducibility. What you have is closer to a compiled binary than source code. You can run it, you can diff it against other binaries, but you cannot, in any meaningful sense, reproduce or extend it from first principles.

This matters because OSS truly depends on the reproducibility claim. "Open weights" borrows the legitimacy of open source (the assumption that scrutiny is possible, that no single actor has a moat, that iteration is democratised). Truly democratised iteration would crack open the training stack and let you generate intelligence from scratch.

Huge kudos to Addie and the team for this :)

]]>Mon, 09 Mar 2026 23:48:33 +0000https://news.ycombinator.com/item?id=47317390oscarmoxonhttps://news.ycombinator.com/item?id=47317390https://news.ycombinator.com/item?id=47317390<![CDATA[New comment by oscarmoxon in "The Blowtorch Theory: A new model for structure formation in the universe"]]>Personally, I love this theory. The thought of natural assembly and selection at the level of Black Holes is alluring. Not sure what The Black Mirror Hypothesis (https://curtjaimungal.substack.com/p/when-you-fall-into-a-bl...) would have to say about this, though.

]]>Wed, 28 May 2025 15:09:12 +0000https://news.ycombinator.com/item?id=44116818oscarmoxonhttps://news.ycombinator.com/item?id=44116818https://news.ycombinator.com/item?id=44116818<![CDATA[New comment by oscarmoxon in "Mullvad Leta"]]>They're also littering the London tube system with ads - there's definitely been a lottery win or a series A.

]]>Wed, 28 May 2025 15:06:54 +0000https://news.ycombinator.com/item?id=44116794oscarmoxonhttps://news.ycombinator.com/item?id=44116794https://news.ycombinator.com/item?id=44116794<![CDATA[New comment by oscarmoxon in "Human"]]>The fact that the machines generated this Wikipedia page for their version of events makes this artwork.

https://claude.ai/public/artifacts/b0e14755-0bd9-4da6-8175-c...

]]>Thu, 15 May 2025 11:00:29 +0000https://news.ycombinator.com/item?id=43993725oscarmoxonhttps://news.ycombinator.com/item?id=43993725https://news.ycombinator.com/item?id=43993725<![CDATA[New comment by oscarmoxon in "Overlay hundred-year-old British maps on modern satellite imagery"]]>The National Library of Scotland has an impressive web tool that allows you to overlay detailed historical Ordnance Survey (OS) maps (dating back to 1841 for this instance) onto modern satellite imagery. The interface lets you adjust transparency and blend between past and present landscapes.

Here is the tool overlaid on Bath.

]]>Sun, 16 Feb 2025 23:26:47 +0000https://news.ycombinator.com/item?id=43072895oscarmoxonhttps://news.ycombinator.com/item?id=43072895https://news.ycombinator.com/item?id=43072895<![CDATA[Overlay hundred-year-old British maps on modern satellite imagery]]>Article URL: https://maps.nls.uk/geo/explore/#zoom=14.9&lat=51.38494&lon=-2.36477&layers=178&b=ESRIWorld&o=100

Comments URL: https://news.ycombinator.com/item?id=43072894

Points: 1

# Comments: 1

]]>Sun, 16 Feb 2025 23:26:47 +0000https://maps.nls.uk/geo/explore/#zoom=14.9&lat=51.38494&lon=-2.36477&layers=178&b=ESRIWorld&o=100oscarmoxonhttps://news.ycombinator.com/item?id=43072894https://news.ycombinator.com/item?id=43072894<![CDATA[New comment by oscarmoxon in "AI, but at What Cost? Breakdown of AI's Carbon Footprint"]]>by this metric, Midjourney produce 1,920,000,000 images/day, that would put it on-par with Google at 8.9bn searches/day!

]]>Tue, 28 Jan 2025 13:34:39 +0000https://news.ycombinator.com/item?id=42852102oscarmoxonhttps://news.ycombinator.com/item?id=42852102https://news.ycombinator.com/item?id=42852102<![CDATA[The Technology Emits the Genius]]>Article URL: https://www.oscarmoxon.com/brains/2024/12/06/the-medium-emits-the-genius.html

Comments URL: https://news.ycombinator.com/item?id=42469328

Points: 2

# Comments: 1

]]>Fri, 20 Dec 2024 08:41:12 +0000https://www.oscarmoxon.com/brains/2024/12/06/the-medium-emits-the-genius.htmloscarmoxonhttps://news.ycombinator.com/item?id=42469328https://news.ycombinator.com/item?id=42469328<![CDATA[Emergent Complexity]]>Article URL: https://www.oscarmoxon.com/hierarchies/2024/10/24/emergence.html

Comments URL: https://news.ycombinator.com/item?id=42381617

Points: 1

# Comments: 0

]]>Tue, 10 Dec 2024 21:17:39 +0000https://www.oscarmoxon.com/hierarchies/2024/10/24/emergence.htmloscarmoxonhttps://news.ycombinator.com/item?id=42381617https://news.ycombinator.com/item?id=42381617