tgrowazay's comments

On: The Waymo World Model

[tgrowazay]

This page crashes my browser.

Vivaldi 7.8.3931.63 on iOS 26.2.1 iPhone 16 pro

On: Impact of Google's Manifest Version 3 (MV3) Update on Ad Blocker Effectiveness

[tgrowazay]

From abstract:

> Through a browser-based experiment conducted across multiple samples of ad-supported websites, we compare the MV3 to MV2 instances of four widely used ad blockers.

> Our results reveal no statistically significant reduction in ad-blocking or anti-tracking effectiveness for MV3 ad blockers compared to their MV2 counterparts, and in some cases, MV3 instances even exhibit slight improvements in blocking trackers.

On: Bitcoin gets a zero price target in wake of Burry warning

[tgrowazay]

One may think it is about quantum computers reversing the keys or something like that, but no.

This piece predicts that BTC will go to zero because it is held my MicroStrategy and is not a good hedge against stock market.

[jfengel → tgrowazay]

The other key piece is AI, which is where the money is going instead. It is newer, so people can hope they're getting in closer to the ground floor.

And unlike Bitcoin, it is producing something. Maybe not what people hope, but it's more than just a "store of value".

Once the speculation money starts to leave, the supply of greater fools will have run out. There will be literally nobody willing to buy it.

On: Claude Opus 4.6

[tgrowazay]

Isn’t that just how many steps at most a reasoning model should do?

On: We built a serverless GPU inference platform with predictable latency

[tgrowazay]

Well, do you have a blog post or we need to ask about each item to get it?

On: Voxtral Transcribe 2

[tgrowazay]

Both AWS and Mistral prices above are per minute of input audio.

[Curiositry → tgrowazay]

If Voxtral can process rapid speech as well as it claims to, an obvious cost optimization would be to speed up normal laconic speech to the maximum speed the model can handle accurately.

On: Students using “humanizer” programs to beat accusations of cheating with AI

[tgrowazay]

Everyone knows about emdashes, but there are so much more!

Here is a wiki article with all common tell-tales of AI writing: https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing

[jurgenaut23 → tgrowazay]

I used em dashes heavily 15 years ago when writing my PhD thesis.

[singpolyma3 → jurgenaut23]

So did every author of classic literature. People who think they can spot AI writing by simple stylistic indicators alone are fooling themselves and hurting real human authors

[A_D_E_P_T → singpolyma3]

It's because LLMs were trained on classic literature that they began to use em-dashes in their now-famous manner.

Seriously, highbrow literature is heavily weighted in their training data. (But the rest is Reddit, etc.) This really explains a lot, I think.

[zeroonetwothree → singpolyma3]

Let’s just say when my coworkers started sending emails with tons of bold and bullet points when they had never done that before I felt pretty justified in assuming they used AI

[SecretDreams → jurgenaut23]

Same, but 5 years ago. Now they're ruined for me lol.

[kbelder → tgrowazay]

My fear is people will actually take that article to heart, and begin accusing people of posting AI simply for using all sorts of completely valid phrases in their writing. None of those AI tells originated with AI.

[AstroBen → tgrowazay]

I saw someone created a skill to weaponize that exact list to humanize the AI's output

There are no clear signs, at least for anyone who cares to hide them

On: Kimi K2.5 Technical Report [pdf]

[tgrowazay]

Just pick up any >240GB VRAM GPU off your local BestBuy to run a quantized version.

> The full Kimi K2.5 model is 630GB and typically requires at least 4× H200 GPUs.

[CamperBob2 → tgrowazay]

You could run the full, unquantized model at high speed with 8 RTX 6000 Blackwell boards.

I don't see a way to put together a decent system of that scale for less than $100K, given RAM and SSD prices. A system with 4x H200s would cost more like $200K.

[ttul → CamperBob2]

That would be quite the space heater, too!

On: Trinity large: An open 400B sparse MoE model

[tgrowazay]

> 2048 Nvidia B300 GPU

With average price of $6/hour that is $12,288/hour for whole cluster.

Times 33 days times 24 hours it comes out to be $9.7MM , assuming no discounts.

That leaves $10.3MM/6 months for salaries, which is 103 employees at $200k/year or 51 employee at $400k/year.

[YetAnotherNick → tgrowazay]

It would likely be something like $4.5/hour for this big cluster.

[1]: https://verda.com/products#B300

[zamadatix → tgrowazay]

It mentions it took 4 models to get there, so would that mean there were additional runs (and other steps/overheads) which were part of the cost separate from just the salaries in that time?

On: Ask HN: Doctors Using AI Workflow Automation – What's the Real Impact?

[tgrowazay]

In my experience, PCPs in places like One Medical are now not shy to ChatGPT symptoms together with a patient. It is kinda funny, but is also helpful and makes sense.