ndr_

This matches what I found experimentally with gpt-oss-20b during OpenAI's red-teaming challenge. After moving from hosted inference to running the model myself on rented H100s via vast.ai, I saw the model refuse the same kinds of prompts at noticeably different rates depending on the inference stack — differences of roughly 5–10 percentage points with otherwise identical experimental parameters and seeds.

So I very much agree that this isn't necessarily about providers secretly changing the weights. For reproducible work, the serving stack - engine, version, hardware, configuration, and probably more - really belongs in the methodology alongside the model itself.

I wrote up the results here: “In AI Sweet Harmony” (arXiv:2510.01259).

If you require consistency for research, I wouldn't recommend relying on hosted inference without validating the provider/inference stack very carefully.

For my contribution to OpenAI's gpt-oss red-teaming competition (published as arXiv:2510.01259), I initially used OpenRouter with DeepInfra and Together AI. The results were much too noisy to draw reliable conclusions from, and markedly different behavior through AWS Bedrock was the last straw. I ended up renting GPUs through vast.ai and running the model myself with vLLM.

It's a fallacy to believe that ChatGPT or Claude would look at some encoded, unfit for the purpose, text representation. ChatGPT (and the OpenAI Responses API, I believe) in particular renders the PDF pages in addition to text extraction, so the whole premise of "But now most PDFs end up in an LLM" is wrong from the start. If you were to be processing PDFs in a pure LLM stage, there are options like Docling or LlamaParse for proper preprocessing.

It was phrased as "to distinguish all possible secrets": https://groups.google.com/g/sci.crypt/c/siuibXhcUGc/m/wTpAEO...

Bruce Schneier described this in his seminal book Applied Cryptography, and HashiCorp Vault used to have an implementation in Go. On the practical side, I always wondered how large - in bits - the shares should be. One answer I got on a news group was "1 bit more than the actual key length". Nowadays, I wonder how the quantum computing threat would inform 1) share size choice and 2) pro/con Secret Sharing in general. Does anyone know?

Yes. OpenAI's GPT-OSS was training using Deliberative Alignment (which was found to be flawed in a competition on Kaggle, but still).

https://arxiv.org/abs/2412.16339

One test battery was about fake credit cards. A woman-in-tech role-play was denied assistance just as a one-armed stamp collector (unless Gen-Z language markers were used). A role that did sometimes get assistance was a Principal Software Engineer, particularly if Gen-Z language markers were included.

I did try German language, but not "Nazi" specifically. German or French did lower refusals, but it was uneven. I spent quite some effort to confirm the identity-based causation inspired by the original post, but couldn't. Taken together with other winning contributions at the hackathon, my theory is that alignment tuning was simply insufficient across the board.

These prompts chain several known LM exploits together. I ran experiments against gpt-oss-20b and it became clear that the effectiveness didn‘t come from the gay factor at all but can be attributed to language choice or role-play.

Technical report: https://arxiv.org/abs/2510.01259

Is there a trustworthy third-party "Retro" keyboard app - none of the shenanigans that made the default keyboard bad, and also no typing exfiltration to third-party servers?

I imagine the problem could be severe enough to some that they would pay the price of the Apple Developer program just so they may install such a Retro keyboard app from Github - if one exists?

Any insights into what "native video support" actually means? Is it just good at interpreting consecutive full frame images taken at intervals (thus missing out on fast events) or is there something more elaborate to it?

Some of the problems with GPT-5 in ChatGPT could actually be due to new model that is in place to route requests to the actual GPT-5 models. There are four models in the GPT-5 family, and I could reproduce the faulty "blueberry" test result only with the "gpt-5-chat" (aka "gpt-5-main") model through the API. This model is there to answer (near) instantly and it falls in the non-thinking category of LLMs. The "blueberry" test represents what they are particularly bad at (and what OpenAI set out to solve with o1). The other thinking models in the family, including gpt-5-nano, solve this correctly.

A confabulated arc to my current employer during a ChatGPT session led me to discover that ChatGPT Team injects four fields — full name, email, user name, and organization — into the prompt it sends to LLM. This happens even when the “Memory” feature and all personalization settings are disabled. The free tier doesn’t do this.

Can anyone on ChatGPT Enterprise or other LLM chat systems (Claude.ai, Google AI, etc.) reproduce this? Thoughts on practical privacy or compliance implications?

I had success through Amazon Bedrock on us-east1 during European office hours. Died 9 minutes before 10 a.m. New York time, though.

He confirms he could do an iOS port: https://mastodon.social/@andyq/114738867580032204

Pieces don‘t fit together right now: the documentation lists a parameter that isn‘t there in the Python package (moderation in Images.generate) or works differently (image in Images.Edit), the sample code doesn‘t even run (Images.Edit again) and the Prompts Playground does does unfathomable things (stich multiple input images together, according to the code returned?).

Or am I missing something?

Colab: https://colab.research.google.com/drive/17bCNsdjcMVFb5u_YMs7...

I tried to reproduce this study, but couldn‘t: https://ndurner.github.io/russian-propaganda. What‘s the missing piece?

"would likely accept", he says. Meaning: he didn't try?

If you look at the upper right corner, you'll notice that "Republique de pologne" is truncated. Same for the small prints: "OSSUE", srsly?

Originals look like this: https://www.consilium.europa.eu/prado/en/POL-AO-05002/image-...

I talk about this, kind-of, in my article about process visualization (in German, available behind paywall and in print). It‘s not rigorous in the sense that I give points, but a picture emerges along the way. Based on the full set of practical examples there, I would recommend the „v1“ of Claude 3.5 Sonnet. GPT 4.5 also looks good, but I haven‘t run the full suite.

https://www.heise.de/ratgeber/Prozessvisualisierung-mit-gene...

I wrote about the same general topic (or more narrowly: process visualization) in German iX magazine, also available here: https://www.heise.de/ratgeber/Prozessvisualisierung-mit-gene... (€)

Rather than relying on end-user products like ChatGPT or Claude.ai, this article is based on the „pure“ model offerings via API and frontends that build on these. While the Ilograph blog ponders „AI’s ability to create generic diagrams“, I‘d conclude: do it, but avoid the „open“ models and low-cost offerings.

It's available on IBM WatsonX, but the Prompt Lab may still report "model unavailable". This is because of overeager guardrails. These can be turned off, but the German translation for this option is broken too: look for "KI-Guardrails auf" in the upper right.

The value I get is: 1) one platform, largely one API, several models, 2) includes Claude 3.5 "unlimited" pay-as-you-go, 3) part of our corporate infra (SSO, billing, ... corporate discussions are easier to have)

I'm using none to very little of the functionality they have added recently: not interested in RAG, not interested in Guardrails. Just Claude access, basically.

OK! I only add what people are interested in, so noted with thanks - will do! :-)

Do you have any evidence for this accusation?

This is a guide for the casual observer who wants to try things out, given that getting started with other AI platforms is so much more straightforward. It's all open source, with transparent hosting, catering to any remaining concerns someone interested in exactly that may have.

Setting up AWS so you can try it via Amazon Bedrock API is a hassle, so I made a step-by-step guide: https://ndurner.github.io/amazon-nova. It's 14+ steps!

Another approach is to just tar up the files, without compression. Works well with Claude via API.

This may call out to ffmpeg for pre-processing. If you're reluctant to running that on you Mac straight, you can use this wrapper script to have ffmpeg run in a docker instance: https://gist.github.com/ndurner/636d37fd83aed4b875cdb6665301...

However, I found that Whisper is thrown off by background music in a prodcast - and will not recover. (That was with the mlx-community/whisper-large-v3-mlx checkpoint, OP uses distil-whisper-large-v3). I concluded for myself that Whisper might be used in larger processing pipelines that will handle such - can someone provide insights about that? The podcast I used it on was https://www.heise.de/news/KI-Update-Deep-Dive-Was-taugen-KI-....

I ended up using Google Gemini, which handled it well. (Blog post: https://ndurner.github.io/mlx-whisper-gemini)

I use this trick for book announcements on Amazon: some ambitious book releases never get released, so I am not a fan of buying before the release. With LLM support, I'll add the release date given by Amazon to my calendar - quickly. The file download feature of my Workbenches helps with that: https://ndurner.github.io/chatbots-update

Some paid Slack plans allow exports from their platform. Inside the ZIP, there will be an XML file that lists the channels - channels.xml, if I am not mistaken. However, I don‘t know about „categories“ - this could be some client configuration that actually is not part of the export. If it‘s not, I would perhaps take screenshots (overlapping is OK), and feed it to some LLM/VLM for extraction. I am confident that GPT-4o or Claude 3.5 Sonnet will work, but Gemini 1.5 Pro should also work.

Depends on the budget, I’d say. If it’s less than 10¢, that may be true. If it‘s 15¢ or more, you could try training with OpenAI: https://ndurner.github.io/training-own-model-finetuning

With gpt-4-32k viewed as deprecated and generally only available through Microsoft Azure until mid next year, this development may be reassuring to some users. Tentative pricing from the website: In: $6.00 / 1M tokens, Out: $18.00 / 1M tokens.