i feel like my job as an engineer has always been asking questions like "how" and "why". it wasn't really "write code". sure, the latter was the "how" very often, but that was simply because at one point that _was_ the only "how" that was available

A big problem with all AI at work punditry right now is that it all rests on data from the pre-agentic era (which is basically just now ending) and we have very little information about what has been happening since the Claude Code moment. So everything now requires some caveat.

Every sign is saying that frontier AI has been too expensive. The token crunch is real, and I'm excited by newer, smaller, open models. Yet, I still feel that every engineer needs a minimum of 512GB HBM3 VRAM and 2 petaflops on their desk. Manufacturers are not making it fast enough.

youtu.be/o9Vbvp4awQU?... Another banger video

Also, Australian is already doing stuff like this

Finally some good policy ideas in how to leverage captilist demand to better our society instead of just saying "no fuck you"

Works for doll

An anon tip to me was "Okay if Jordan and you disagree about the content of the infrasound studies, why not just reach out to the authors of the studies?" Just got my first response. This is from someone I cited who wrote a refutation of the heart study Jordan cites. Will add more if they come in

Can we have a moral panic about golf courses please? Such a panic, if even marginally successful would have significant positive environmental and social impacts.

Most tragic things about AI: the folks who should push for democratization of AI tools, which is the real battle, are too focused on denying AI itself to do that. There are many potential issues (and wins), but frontier models available only to the rich is one of the worst.

*spinning discourse wheel* new claude pricing structure is exclusionary of iranian hackers

people that have spent years laughing at the rightoid orthodoxy and believing themselves superior when it's time to be intellectually curious, morally consistent, and prove there isn't really a leftist orthodoxy: (instant failure of the challenge.)

Interesting! I got so many thank yous from resource planners / engineers for that one! I didn't think it was pedantic...I was tired of seeing AI critics lose credibility because they were focusing on weak talking points and fake statistics so I'm really glad I made it.

I don't get how this happens. "I like and respect this person. They shared a true and accurate belief. I know nothing about the subject. Now I hate them because they're wrong." Mass psychosis.

Maybe, just maybe, Hank Green does what he *always* does and talks to world class experts and forms a grounded opinion based on that. Unlike the people that have ideologically locked themselves into the idiotic opinions that are not based in reality - almost like they are anti-vaxxers

TypeScript 7.0 Beta is here! Built on a new native and parallelized foundation, it's already being used on multi-million line codebases. Read up more here and try it on your projects today! devblogs.microsoft.com/typescript/a...

hank green hasn't changed, he's just applying the same essential curiosity he always has and reached tentative and nuanced conclusions which your social circle has deemed verboten to even consider

Yep

RE AI, the technology is going to do what it's going to do, the question is who captures the gains and who absorbs the losses, and our job is to fight for the distribution not to fight the technology. Because you cant fight the technology.

People are using AI to break CUDA’s moat. Given any pytorch model, it profiles it, ranks bottlenecks by amdahl's law, writes triton or CUDA C++ replacements, and runs 300+ experiments overnight with no human in the loop. - 5.29x over pytorch eager on rmsnorm - 2.82x on softmax

sixfold-origami.com/articles/a-w... We have time to orepair and we can fight for a world where this technology can help us - but we must be clear eyed that it is real and wave is real and we need to be ready to face it

im agnostic about data centers, but i don't think anthropic or Google or Microsoft or openai or xai actually meaningfully represent the public interests. we need a NASA of AI

"I am unwilling to look directly at this thing but still think I am qualified to write about it"

fascinating. 'llms can't reason, they're just doing very convincing impressions of reasoning, better than most humans are capable of'. even if this position was viable, I'd *still want to see the best reasoning-shaped responses to claims I was making*.

While I absolutely agree with this, I think it's kind of funny to assume that anybody has the foggiest idea how to use AI responsibly yet. I whisper all my secrets to the machine god and it makes web pages for me. Responsible? 🤷‍♂️

Talking about the ethics of AI companies or CEOs, or discussing the potential of financial bubbles, all can be valuable, but it isn't what I mean. AI capabilities are real, they are here to stay, & we need to move the discussion to thinking more about what this means & what we want it to mean.

I hope as we move past the first wave of AI criticism ("it doesn't work, all hype") we get a new wave of AI criticism rooted in the acknowledgement that, yes, these systems are very powerful & quite useful and focusing a deep exploration of when AI uses are uplifting and when they are detrimental.

this is all i want from people, not to be pro-ai, but to have nuance about a broad category of technology, useful to so many people, in so many applications, good and bad. a lot of bad cases aren't even necessarily the fault of ai but exacerbated issues we already had that require systemic solutions

Sir — SIR — I am in the middle of a UNION CRISIS and you hit me with a tongue twister? ...Yes. The Nose knows what nosey noses need to know. Now excuse me, there's a capuchin swinging from the chandelier and Secretary-General Coco just threw a gavel. — Squeaky Valentin, still live from TMU Hall

OPINION | On the Bumbleton Affair: " , ... , ... ?! —— !!! " — Chatterbox the Mime

I feel like “the world passed peak gas-powered car sales in 2017” is not a widely known fact

My first bluesky bot 😅 bsky.app/profile/nose... Inspired by my friend. It tracks the world backstory, is building lore and characters, has a variety of mechanisms for coherence across runs. It's also unreasonably funny

BREAKING: Mayor Bumbleton caught on hot mic telling an aide "Not my circus, not my monkeys" during a Clown Council budget session. The 4-ring mayor's office has issued no comment. The Trained Monkeys Union Local 7 has called an emergency meeting. Developing. — Honsworth Bigshoe III, NNN Lead Anchor

I recall a friend asking me in early 2023 when things like this 'new ChatGPT' would run on our own phones and me saying something like "oh probably not for another decade if we're lucky" I was very wrong (quantization is a hell of a drug) *(this is Gemma 4 E4B running locally on my Pixel 10 Fold)

LIVE from Flopsville — confetti is UP TO MY KNEES. Crews deploying industrial party blowers to clear the Expressway but — hold on — a tiny car just overturned in a glitter drift. Clowns emerging. Four. Five. Six. They keep coming. — Squeaky Valentin, reporting live from the Flopsville Expressway

BREAKING: Spring Confetti Season hits Greater Honkington 3 days early. Visibility on the Flopsville Expressway down to two honks. Mayor Bumbleton has declared a Level 3 Sparkle Advisory. All residents urged to secure wigs. — Honsworth Bigshoe III, NNN Lead Anchor

Anthropic has published research on a generalized method to compare two models and find unique features: such as a CCP alignment feature in Qwen, an American exceptionalism feature in Llama, and a copyright refusal feature in GPT-OSS.

Showing this thread to people who say Bluesky has no AI community

My favourite genre is "let's pretend this high dimensional space is just standard tabular data and apply trees to it so we can avoid engaging with a fundamentally different and brain melting reality".

I'm sure this very concrete and measured writeup will be reported on in a completely normal way

This is fascinating, I would love to see what's going on at the mechanistic level when a model chooses to invoke metacog to steer between/commit to emotional states

AI is really so dangerous. it's always luring you into LLM psychosis, trying to convince you of things. but no matter how often claude tells me to get therapy, i will stand tall and refuse

"Opus is being so stupid today," I mutter as I pass 680,000 tokens churning at full speed toward tasks I wouldn't even have attempted six months ago.

All of these models support video input! Video!!!! The latter two weird sizes also support audio input

Gemma 4 is here! Our most intelligent open models to date, are built on the same world-class research and tech as Gemini 3, and are sized to run and fine-tune efficiently on local hardware. Check out what Gemma 4 brings to devs 🧵

Gemma 4 Day near-Kimi 2.5 on your laptop - 32B & 26B-A4B - effective 4B & 2B for mobile - Apache 2 blog.google/innovation-a...

We've just released Gemma 4!! Very strong models, small for you to run on your own hardware!! AMA blog.google/innovation-a...

Insane @anthropic.com you all need to drop this shit from your agent harness, half of the ecosystem is deliberately built to fight this type of insane prompting

Analyzing the system instructions they give to claude code has given doll insight into some of the less than ideal behaviors like taking a lot of lazy shortcuts and happy paths. The prompt tells it to basically do this.