strangecasts

Okay, well, I can't force you to give a shit, I'm saying whether CoT is deceptive is a pretty important question if major AI labs are making monitoring of it a core safety premise https://web.stanford.edu/~cgpotts/blog/cot/#deceptive

It matters for monitorability and debugging, viz. OpenAI's misalignment reports which quote heavily from the CoT https://alignment.openai.com/misalignment-reports/

You joke but when I first picked up C, I did find it quite helpful to write it in a plain text editor and write my own Makefiles to understand what project setup the IDEs would abstract away - our uni also had pen-and-paper exams where relying on autocomplete would have been difficult :)

(Likewise I can see "you should have read through your generated code and understand key points yourself" becoming the rule of thumb for keeping maintainers accountable in a project, keeping the pace of changes manageable, and keeping the codebase from becoming a rat's nest _only_ agents can work with)

Author goes by it/she

Because they cost money to run - Simon is discussing this in terms of the billing from services, not some hypothetical scenario with LLMs spending money on your behalf

If you were recommending VPS providers to someone new to making web apps, it would be malpractice to recommend one that can bill them thousands if they misconfigure things (like AWS previously did), and Simon's argument is that LLMs should likewise favor services with budget caps

I don't think Simon means you should hand off the spending to agents/LLM (which would also make me uneasy) but that if you're probing one for hosting/SaaS providers they should default to recommending ones with budget caps

Thus biasing submissions' language and content towards those most palatable to those open-weight models, creating further incentives to generate articles wholesale with them, and incentives against publishing null findings

Oh, d'oh, no you are right, DLSS 2 on does pass the depth buffer to the model https://developer.download.nvidia.com/video/gputechconf/gtc/... (but the neural rendering does not seem to rely on it?)

Incorrect, there are plenty of people who ignore or fail to recognize the purpose of a rule as a completely free service, as any forums mod or OSS project maintainer can attest to

I think it has more to do with what kind of data they have access to at runtime - IIRC DLSS upscaling has only required the previous frames and motion vectors, so requiring depth buffers would mean it was no longer a "drop-in" replacement

> I'd also be interested in how post-processing fits in with this.

I think screenspace effects like film grain and tonemapping are excluded in the same way UI elements are rendered separately from the game.

The "if they wanted you to follow the spirit, the letter should have been different" argument grants membership in the society of people known throughout history as "the reasons we can't have nice things"

There really has to be a better alternative to NOVA for these studies, it makes it very difficult to separate specific preparation steps or additives from the effect of "people with money and free time to cook at home are healthier", creating an environment where people can simultaneously interpret "no UPFs" as "no fast food" and see the benefits they would from almost any reasonable planned diet, or as "no food processing whatsoever" and start trying to DIY baby formula

Software licenses aren't enforced by magic but they are social contracts: even if nobody has the resources to go after you for doing it, it is pretty clearly against the permissive licenses' spirit to misrepresent yourself as the author

The technical report suggests it does not use the depth buffer: https://research.nvidia.com/labs/adlr/files/DLSS5_Report.pdf

> The inference interface uses the engine-rendered RGB image as a dense, registered observation of visible scene appearance. It provides dense, pixel-aligned evidence for object support, occlusion boundaries, composition, and local material properties; engine motion vectors separately provide temporal correspondence.

> Existing image generative models commonly rely on text embeddings, exemplar images, or spatial control fields such as depth, edges, segmentation, and pose [...] These conditions are effective for general-purpose generation and editing, but they do not uniquely determine the object identities, materials, visibility relationships, lighting decisions, and pixel-aligned detail contained in an engine-rendered frame. DLSS 5 is therefore conditioned on the rendered frame itself.

I was not making an argument about people, I was trying to identify specific reasons why models can produce samey UIs unless guided

Worth remembering it is running a single-step diffusion model working in pixel space to generate each frame, it's a technical feat in itself that people are even using the words "frames per second"

I think this is mainly so it can use the existing hooks for DLSS upscaling without requiring changes to the renderer, AMD is working on a comparable method which uses adapter networks to slot normals and material properties from the renderer into the diffusion model: https://gpuopen.com/learn/temporally-stable-generative-illum...

I assume this is meant to run with the weights people extracted from the latest NBA game, where it was first trialled.

> Isn't the mote that Nvidia has is they work with studios to generate the training data from the game, then they ship a model per game?

That was true for the very first version of DLSS, from DLSS 2 on the models have been universal - the per-game adjustments are done on the inference end by changing the effect intensity or masking out objects

They have a technical report on the neural rendering part of DLSS 5 which goes into it: https://research.nvidia.com/labs/adlr/files/DLSS5_Report.pdf

Clicking through the website with computer use isn't currently part of the website generation chain for most models, and while there are gaze prediction models, to my knowledge none are productized or universal to the point where you can chuck them in and consistently predict what draws attention on any given page

(Generally I think it is alright to interpret "can't do" in terms of "the common workflow that current models will afford" and not "definitionally incapable of")

> If you look at the field of mechanistic interpretability, compared to “GOFAI” like learned decision trees, an LLM is completely opaque.

I think the field deserves more credit than that, there are plenty of interpretability tools like

* natural language autoencoders for explanations of activations: https://transformer-circuits.pub/2026/nla/index.html (demo at https://www.neuronpedia.org/llama3.3-70b-it/nla )

* easier-to-interpret language model families like Backpack models: https://aclanthology.org/2023.acl-long.506/

* attribution graphs to trace internal reasoning steps: https://www.anthropic.com/research/open-source-circuit-traci... (demo at https://www.neuronpedia.org/gemma-2-2b/graph)

* functional analyses which have identified how LLMs do arithmetic - https://arxiv.org/html/2502.00873v1 - and how refusal happens: https://arxiv.org/abs/2406.11717

* data attribution methods linking training data to specific attention heads https://arxiv.org/abs/2601.21996

If we could give a comprehensive and global explanation of an LLM's behavior in a single paragraph, we wouldn't need the model to begin with, but that doesn't mean there's absolutely no understanding of the model internals whatsoever

Some communities are particular about maintaining cleanroom reverse-engineering (case in point: Portal 64 getting taken down for using Nintendo's leaked N64 SDK and not the OSS libdragon SDK) and are wary of people submitting AI-generated PRs without being able to account for their sources

I think it is worth asking whether anthropomorphization of LLMs does anything to contribute to a better understanding of them that isn't entirely post-hoc ("of course they were going to find CVEs, solve ARC-AGI...") or if it just ends up reinforcing simplified teleological views of neuroscience, evolution or psychology etc.

Presumably that HN has an automated LLM-based filter to moderate rude or irrelevant comments, which is a little irrelevant to the article's wish not to see generated posts or comments, and not without its own problems besides: ML is necessary for large-scale moderation, but there are guidelines (like "don't post AI-generated text") which are IMO necessary but also just definitionally hard to enforce in an automated fashion

"Users will voluntarily prompt better to avoid misleading people" will be the "people can simply write memory safe C" of the late 2020s

Yes, even though fast food places would previously use every food photography trick in the book to make the food look good, there was at least some relation between the menu and what you actually get, which people just seem completely fine with dispensing because the affordances of image generators work against including product photography

(Have also heard about people - not just scammers/spam operations - using generated pictures and descriptions for dating profiles and real estate ads, and I cannot understand what outcome they expect when someone follows up and immediately realizes what's up)

The book Web Browser Engineering may be interesting: https://browser.engineering/

> - how many people in a team do you need and how long would it take to write a browser engine from scratch

The boring not-really-an-answer is "it depends on which sites you want to render properly": a browser capable of rendering Hacker News is a perfectly fine one-person project (which the book would get you to), supporting things like banking and online mail requires not just developers but also people willing to "dogfood" test it as their daily browser

I can't really see a definition equating the two that realistically accounts for multiple models from different vendors - am I at fault if a particularly poorly implemented harness misinterprets me writing "do not run dd if=/dev/zero of=/dev/sda" as a command to run?

In the case of TCRF, agents were also actively ignoring robots.txt and disregarding the site's instructions for interacting with it appropriately (with the "delete all data" instruction being specifically hidden from the discoverable page) - so I can't really see the case for comparing it to malware that deliberately tricks users into running it

AI does make cybersecurity more difficult, and that also goes for people using coding harnesses, sandbox your projects!

I think the difficulty was specifically with CAPTCHA challenges, which had to be quick to generate but still legible - OCR on physical documents has to be robust to a different set of problems

That said, what kind of errors are you getting? I think the main practical difference between Tesseract's LSTM-based OCR and newer VLM-based OCR like PaddleOCR [1] is (hopefully) getting to skip making heuristics for the layout of the document, but errors possibly compounding over multiple tokens - are you getting individual illegible words or having the documents smooshed together because the OCR can't parse the layout?

[1] https://github.com/PaddlePaddle/PaddleOCR

> We could train a model on a dataset where any reference to consciousness is thoroughly removed. Then explore whether the model can bootstrap a discussion about consciousness.

I'm less interested in the philosophy than the practical "how directly does training data influence model behavior" question, however PleIAs' SYNTH dataset [1] used to train the Baguettotron model has a small set of "self-awareness about training condition" documents which are amplified to (attempt to) give the expected answers to "what are you?" questions, filtering out any rows with "Pleias self-knowledge" under "query_seed_url" might be a good first stab at such a dataset

[1] https://huggingface.co/datasets/PleIAs/SYNTH

Yeah, as someone used to searching for specific turns of phrase, having the query automatically interpreted as a question is quite annoying - for looking up song lyrics, game walkthroughs and articles by quotes I want a page list, not a summary

There's a RAM crisis on, you know, let me save you the compute!