hendersoon

I setup 8 day age gates on everything back in May. This is inobtrusive and pretty safe.

STT is definitely better on my iPhone 15 pro which doesn’t get the large model. It’s usable now but not up to the real deal like parakeet with a LM cleaning up the output after.

Yes, that was Google’s justification for killing MV2. It isn’t wrong, of course, but I insist on defining my own security posture and I prioritize adblocking.

Brave and Vivaldi built their own blockers, they aren’t addons.

What is the purpose of this? Just a hard cutoff below the actual context window? You could set that in your harness anyway.

This was done because they were charging the 100%/50% premium for long-context over 272k to Codex users, even though they set 372k as the default. It cost people a LOT of money.

Parakeet is incredibly fast and accurate even on CPU, and it supports streaming now also in TDT3.

OK, but how does it compare against Parakeet TDT2 (english), TDT3 (many languages), and Parakeet TDT3 Streaming? And what about whisper large?

Atlassian was extremely foolish. Obviously you want to fire someone who makes unprofessional comments like that on the company Slack, but you don't explicitly fire them for that, you find some other defensible reason to do it.

If your prompt has to do with those areas, yes. I haven't seen a single refusal yet.

Reportedly the biology guiderails are particularly strict.

Must be to classify/moderate images for social media. They're pretty good at that. I can't imagine what else you'd want to use Gemini models for, certainly not coding.

All I can say is I will never vote for any politician who votes for any form of this. Even if the bill fails to pass, they will never, ever get my vote.

No actual proof of any kind. Obviously a petulant attack on Anthropic.

Claude code basically fixes MCP context usage with tool search, so MCPs are only loaded into context when actually used. Unfortunately codex doesn't support that functionality.

Until that happy day arrives I run every required MCP with mcpc.

[1] https://github.com/apify/mcpc

There's a big difference between "Yo GPT, copy this webpage for me in a different voice" and blaming LMs wholesale for being plagiarism. The former is of course a problem. The latter warrants a much more nuanced discussion about learning and generalization.

Oh yeah? Still think so?

https://github.blog/news-insights/company-news/github-copilo...

That was exactly my question; why fork? I assumed the project owners were preserving their business model, but if you allow it to be opened-up and turn off the cloud features, I see no reason to fork.

As of this week, the "request" gravy train is over. Github copilot now meters tokens not requests.

The forced telemetry and blocking local user accounts is particularly egregious. I've been running Shutup10 for over a decade now, disabling their telemetry after every update. It's outrageous that they force mt to do this.

And local accounts, all the methods required following a guide, using a hidden hotkey like shift-F10 and typing in an obscure command. Nobody did this by accident, or was coerced into it. These were sophisticated users who did not want to login to a MS account. Microsoft not only didn't care about their users' preferences, they actively fought us. It's downright offensive.

And like a battered spouse, we lived with it, we took the hits, because we wanted to play videogames. Well, now Windows is no longer required, thanks to Valve's work on WINE. And suddenly they're apologetic and conciliatory. No thanks, Microsoft.

Executive Order 12333, referenced in the OpenAI agreement, has been widely used by the NSA for warrantless mass surveillance of American citizens without judicial scrutiny. It does prohibit targeting specific American citizens.

This EO is far more permissive than the FISA act from congress which explictly forbids bulk collection.

In naming EO 12333, this agreement explicitly allows using OpenAI services to conduct mass surveillance on American citizens, so long as that data is gathered abroad and specific citizens were not targeted when it was gathered.

But it's bulk collection, so your data is in there.

The -p flag should be fine, so long as you don't use their oauth in a third-party tool. Gemini also supports A2A for this sort of thing.

Yes, I also use Handy. It supports local transcription via Nvidia Parakeet TDT2, which is extremely fast and accurate. I also use gemini 2.5 flash lite for post-processing via the free AI studio API (post-processing is optional and can also use a locally-hosted LM).

Could well be running on Google TPUs.

Yes, this was a defensive move from Nvidia.

My understanding is Groq failed to deploy their second-gen chips on time, which caused their stock to deflate.

Groq's primary advantage over Cerebras and SambaNova, as I see it, is they don't fabricate on TSMC. That's attractive to Nvidia, who doesn't want to give up any of their datacenter GPU allocation.

So with support for OCI container images, does this mean I can run docker images as LXCs natively in proxmox? I guess it's an entirely manual process, no mature orchestration like portainer or even docker-compose, no easy upgrades, manually setting up bind mounts, etc. It would be a nice first step.

Chrome can do that too on desktop, and on iOS Chrome can't run any extensions at all. Safari web extensions have been around since iOS15, so several years now.

It's just another declarative adblocker, as that is all Safari (and now Chrome) allows. There's vanishingly little room for differentiation in this space.

MCP is convenient and the context pollution issue is easily solved by running them in subagents. The real miss here was not doing that from the start.

Well, stdio security issues when not sandboxed are another huge miss, although that's a bit of a derail.

Google approached this the right way. No, not with "ai mode", that sucks. With the Chrome dev tools MCP. You allow AI to control the browser if the user opts-in and sets it up.

Exactly right, the OCR isn't the interesting part. 10x context compression is potentially huge. (With caveats, at only ~97% accuracy, so not appropriate for everything.)