Jackson__

It does really take a load off my mind, knowing that he's completely out of binance now, and instead they have an independent CEO like... his wife. Yep. Can't get more independent than that.

>To be done right, such vision benchmark should consider the peculiarities of the vision tower

As per my previous comment, I used the _exact same weights_ for the comparison. The vision tower is always kept as an unquantized BF16 file for GGUF, as I believe is the default.

>Second of all, the "find coordinates" of something is super difficult task of any model

It is in fact not (anymore), and most recent VLMs I tested have been standardized to point and bbox in a relative 0..1000 coordinate system with decent enough accuracy.

And finally, none of this excuses that Strata performs so much worse with the same weights. Which does kinda leave me confused as to what the point of this reply is.

I've just tested Strata on a simple 50 image vision benchmark. The task is to output the exact coordinates of a requested object. The result via Strata had a median error distance of 154.8 pixels, avg of 168.8. Running the exact same GGUF and vision adapter weights on llama.cpp gives me a median error of 46.5, avg 81.4.

To put that into perspective, here are some more numbers from other models via llama.cpp:

Median/Average

Qwen 3.5 9B BF16: 46.5 / 193.3

Qwen 3.6 35B Q4 K XL: 38.4 / 76.4

Qwen 3.5 122B Q3 K M: 32.9 / 68.6

The difference in vision performance is as large as the jump from a 9B model to a 35B model. All tests were performed at temp=0.

I have done no further testing, as these results line up perfectly with my expectations.

Parallel construction has been a thing for a very long time to get around this.

Judging by the performance of the shown examples, the quality is closer to pre-2022 Photoshop content aware fill than actual 10B models.

I think it is safe to say this is pretty far from a "scientific" work.

If you can't trust them to act ethically on the small scale, why would you expect that to turn around once it gets to a larger much more important scale?

How many government sanctioned school bombings does it take for them to quit working with said government? For now we know that number is somewhere between infinity and 1.

No one inside SpaceX, except for Elon Musk himself? https://arstechnica.com/space/2025/05/spacex-pushed-sniper-t...

Oh what is that, the most "ethical" AI company on the planet making deals with literal democracy undermining fascists?

I'm starting to think the problem with "ethical" AI was always that no company could ever act ethically in the long term. They are and always will be a cancer to society and AI will only serve to amplify this further.

Alternative take: Because no designers are getting paid to move "rm" to "fileops rm" or otherwise between releases.

Yet it just so happens OAI donated millions[0] to the trump admin in the past. And they were immediately there to pick up the slack.

Call me a conspiracy theorist, but this sounds like classic quid pro quo. I would not be surprised if the ousting of anthropic was in part caused by these donations.

[0]https://www.nytimes.com/2024/12/13/technology/openai-sam-alt...

https://finance.yahoo.com/news/openai-exec-becomes-top-trump...

Hell, I would have thought it likely that anthropic was doing the same thing. Of course that was proven wrong, but for OAI I wouldn't even be guessing. This has always been what sama does.

As someone who is new to the whole google cloud ecosystem, the amount of dark patterns they employ are absolutely shocking. Just off the top of my head:

1. You never know how much a single API request will cost or did cost for the gemini api

2. It takes anywhere between 12-24 hours to tell you how much they will charge you for past aggregate requests

3. No simple way to set limits on payment anywhere in google cloud

4. Either they are charging for the batch api before even returning a result, or their "minimal" thinking mode is burning through 15k tokens for a simple image description task with <200 output tokens. I have no way of knowing which of the two it is. The tokens in the UI are not adding up to the costs, so I can only assume its the first.

5. Incomplete batch requests can't be retrieved if they expire, despite being charged.

6. A truly labyrinthine ui experience that makes modern gacha game developers blush

All I have learned here is to never, ever use a google product.

As your local vision nut, their claims about "SOTA" vision are absolutely BS in my tests.

Sure it's SOTA at standard vision benchmarks. But on tasks that require proper image understanding, see for example BabyVision[0] it appears very much lacking compared to Gemini 3 Pro.

[0] https://arxiv.org/html/2601.06521v1

It is literally not even a vision model.

It's not just any pile of melting glaciers, its a pile of melting glaciers that looks really big in mercator projection!!

Me wonders how many millions of peasants he is ready to throw into the wood chipper to pay for this. Just kidding, of course I know it's all of them.

>6. Once validated, the code is good for 1 year (or 6 months or 3 months, adjust based on how stringent you want to make it) - then it expires and a new one must be purchased.

> 7. A separate token is required for each website/each account.

I propose instead:

A single code valid for 10 packets sent to a single IP address, or 30 seconds, whichever expires first.

See also "The only moral abortion is my abortion" for the complete opposite. Where people fail to develop empathy even after it has affected them.

https://joycearthur.com/abortion/the-only-moral-abortion-is-...

Funny that, their front page demo has a mistake. For the waves simulation, the user asks:

>- The UI should be calming and realistic.

Yet what it did is make a sleek frosted glass UI with rounded edges. What it should have done is call a wellness check on the user on suspicion of a co2 leak leading to delirium.

This once again shows that idiocracy was an overly optimistic movie.

It exceedingly appears as if the stock market is becoming little more than the means to liquidate the commoners savings to further prop up the rich.

So they spent all of their R&D to copy deepseek, leaving none for the singular novel added feature: vision.

To quote the hf page:

>Behind vision-first models in multimodal tasks: Mistral Large 3 can lag behind models optimized for vision tasks and use cases.

I'm referring to the plane loads of hardware they shipped in during the tariff panic.

It's not like apple is going to eat the cost of this for either though. As soon as their tariff supply runs out they'll price hike like everyone else.

Sadly pipewire still has issues properly delivering audio on my system with all core load of ~50%. It makes media consumption on a linux pc simply impossible for me, even a 13 year old thinkpad running windows is better in that regard.

what are you hinting towards?

From the outside, it always looked like they gave LeCun just barely enough compute for small scale experiments. They'd publish a promising new paper, show it works at a small scale, then not use it at all for any of their large AI runs.

I would have loved to see a VLM utilizing JEPA for example, but it simply never happened.

Eh, this is way overblown IMO. The product page claims this is for training, and as long as you crank your batch size high enough you will not run into memory bandwidth constraints.

I've finetuned diffusion models streaming from an SSD without noticeable speed penalty at high enough batchsize.

You will never guess what the upcoming gemini 3 has been benchmaxxed upon ;)

Yeah, the blog post took a bit of a weird turn there + the AI slop headers are also quite off putting. I know this may be conspiracy territory, but it does make me wonder whether this is guerilla marketing to ensure big E meets the sale goals for his $1T payment.