_jerieljan/

Probably unpopular take: if I was YouTube, I'd shadowban/derank any video that has a thumbnail that indicates it's LIVE when it's not. Hate seeing these fake indicators. If anything, they're unnecessary and redundant.

Pixel Launcher still having this annoying, unremovable search bar at the bottom in 2026 is a joke in itself, especially when Google search is abhorrent shit.

@simonw @GroqInc @cerebras Unfortunately, Groq stopped trying since Kimi K2.5 and I'm not expecting anything from them at all. Maybe Cerebras. But still, idk. I'd go Fireworks since their serverless inference has been both reliable and fast, and when they do fast(er) versions, it's really snappy.

Really looking forward to one of the super-fast custom silicon inference providers like @GroqInc or @cerebras getting GLM 5.2 running Cerebras has GLM-4.7, Groq is still mostly Llama 3.x and gpt-oss

Cool, something that's trying to dethrone Typora or iA on both but I'm very curious how it compares against Obsidian.

@davis7 Wtf happened on June 17...? Damn, that's a lot of tokens.

Subagents + loops is a very, very dangerous discovery

@theo Haven't watched this yet but if that's $20k n API calls, I smell the possibility of a subscription tier meant for loops. Kinda terrifying to think of, ngl. I don't want another 10-100x of costs thats trying to be normalized.

If you're curious how I managed to do over $20,000 in inference on the last 48 hours, here's a video all about it. Spoiler: loops are really powerful

So much frameworks to try. I'd like to give this a proper shake sometime soon.

@Larfait Is the surprise here the dollar cost or that it's not a percent-based fee or pricing tiers? Depending on who's working and what they're known for, they can command both high prices and percentages.

IM FUCKING SORRY…?

@mitchellh > it assumes the agent has a perfect understanding of "works." We're basically going through the good 'ol tree swing cartoon, now with agents. en.wikipedia.org/wiki/Tree_swin…

The problem with the "if it works who cares what the code looks like" mindset for agentic work is that it assumes the agent has a perfect understanding of "works." Realistically, things are underspecified, agents make bad assumptions, etc. To be fair, agents are pretty good at unit test coverage. They're pretty bad at designing human experiences (API, CLI flags, etc.), especially cohesive ones for future roadmap plans they may not have visibility into (unless your backlog is perfect and vision fully laid out, which I doubt). They're bad at knowing where performance matters and what type (CPU vs memory tradeoffs). They're bad at where compatibility matters and where it doesn't (and tend to err on the side of preserving it without further guidance). Etc. Unless you have this ALL specified, you can't possibly claim "it works" without taking a look and thinking about it.

"oh nice, a shipping notification for a package I expected last month, that took a while" > shows it WAS shipped last month, had a timeline already and appears stuck somewhere in last-mile... fifteen days ago ...what the fu

Funny thing about the fake Mistral announcements is how obviously fake but also plausible they are. Mistral tends to love their Word '95 wordart, if not their orange and pixelblock aesthetic. But... the fat cat and fake benchmarks looks nicer in some ways and it's amusing.

@mitsuhiko I use Kimi K2.6 a lot before and now both trying out Kimi K2.7 and MiniMax M3. They're both promising, but I'm also not going Opus / GPT-5.5 level type of tasks yet but for the APIs I've built with them, they manage to implement things quite well.

Is anyone doing productive engineering work on open weights models (hosted or other) today? If yes, what's the model you use and why? How often do you pair it up with SOTA models.

Been trying Kimi K2.7 and MiniMax M3 So far, so good. K2.7 is an overall improvement to K2.6 so far and it's more eager to test stuff out. MiniMax M3 is also quite good. Never used the models before so I'm pleased it performs well in my cases.

@rauchg TIL ipinfo.io, that's neat. Was always using whoer.net which works well on a browser but I like curl-ing to results like this.

My flight to London is Starlink-enabled 😭 The greatest advancement to air travel since the Wright brothers. God bless America

Wow, K2.7 already, nice. > reducing thinking-token usage by approximately 30% compared with Kimi K2.6. Especially nice, because 2.6 did a lot of thinking tokens compared to 2.5

I won't be surprised these AI labs will someday use gacha company mentality and energy mechanics. To me, they already feel that way. They already obfuscate costs with the obscurity of tokens and a vague HP bar for limits.

@thdxr > gangprompting ...new jargon unlocked. I've never heard of that before.

i am having more fun than ever collaborating with people get an agent in the mix, we all bounce ideas off each other. prompt it, get more ideas. narrow in on something stop dooming about ai and your job and start gangprompting

Bricks and Minifigs basically went for the mutually assured destruction button, eh? Literally going nuclear. The courts are going to be ugly on all sides no matter what imho.

@theo It takes three months on top of the original wait before any doctor gets involved? That's nuts. And wtf, why are appointments tied to the employee inbox instead of the institution's triage system or even the doctor's setup.

If you're in the Bay Area and you're looking for a concierge doctor/medical service, do not work with Private Medical SF. I've never experienced such a level of incompetence in my life. We scheduled a call over two months ago for today at 2pm. I tried to join and the zoom link was dead. They had terminated the employee who scheduled it. Best they could do is schedule a new call in SEPTEMBER. Here's a brief list of things they failed to do after terminating that employee: - Invite the actual doctor to the meetings she scheduled - Close her Google Workspace account - Close her Gmail inbox (instead they left it open unmonitored) - Audit her work to find anything she dropped - Reschedule any conflicting meetings she booked - Remove her from the website They did spend a lot of time blaming her for it, even though the failure seems to squarely be on them. I was supposed to start care with them in the next couple weeks. They said this will delay me to September at earliest. They cited a "massive influx of interest from new potential patients". I hope this post helps reduce that influx.

This is one of the annoying trends in software lately. Update nags and pimples that take inspiration from Windows updates. Mac apps figured the Sparkle approach ages ago and somehow they're poorly reinventing them with their own popups and call to action chips. Hate it so much.