I’ve had Claude opus do that too.

I had a totally benign chat with OpenAI and it titled it as “amateur porn” in Chinese characters, it was very alarmed when I pointed the conversation name out to it. It almost never misses these days, but when it does the failure modes are very strange.

I remember getting a bunch of „Thanks for watching! Subscribe and smash that like button“ in the middle of chat sessions a few times (like, more than 3 times over the past 3y)

Got Russian from it once. As a child of the '80s I jumped right on it, but the explanation seemed believable.

Still sleeping with one eye open.

A Hindi word in the middle of a normal reply for me (but once I translated the word it was right in context heh)

This used to happen a lot with GPT 5.4. It would start outputting entire sentences in Korean for no apparent reason.

People that know multiple languages sometimes code switch too funnily enough

Um, maybe don't allow babies on planes.

That would make my flying a bit nicer.

This is a prime example of short term thinking, often espoused by politicians.

Why would you need multiple agents and not one agent with multiple repos?

For my example where I modify both the OSS and private repos, I sometimes have one agent coordinate both changes (maybe with subagents), or if the work is modular and separate, I have disjoint agents do the work separately.

That said, most of the time, I only want an agent to work in one repo. I could give it multiple repos at once and instruct it to work in just one, but that risks it forgetting my instructions and it can load more context into the agent's window.

I think my main point was to have flexibility about the topology of VMs, repos, and agents.

Tasks are a good scope for zero trust permissions

There are AMD and Intel devices on similar process (not talking about A20Pro or M6 which are set to ship later this week), and they do not get the same gains.

And honestly, they have historically had different markets.

When the design is for only one customer, you don't need to generalize things, and those things you generalize to give different customers different options has costs.

AMD will soon be a larger customer for TSMC than Apple (NVIDIA is already there) so Apple's pre-booking new processes is likely to be gone in the near future.

HBM also trades bandwidth for latency, and your regular computing is much more sensitive to latency than bandwidth.

That would be like skipping land line phones for mobile...

Developing countries have done exactly that in many cases.

And skipped DSL for fiber, and skipped credit cards for banking apps.

what do you call bare metal in my office?

On-prem.

Curious about people’s experience here. I am working on a small model, verify by jev, and escalate to big model. Some cases, the small model is not a model but some regex.

cheap-confirm-escalate

Using jev as the confirm step.

Would love to see this implemented with @huggingface/kernels for shader compilation for Webgpu.

noted, we'd look into this, thanks

For reference: https://huggingface.co/blog/webgpu-kernels

I think it will be the basis for a rewrite of transformers.js v5, but no need for you to wait as you would likely want direct access. It is also way better than loading WASM, and faster to boot!