lysecret

I don’t get it we are paying at least hundreds or maybe thousands per month on ai costs. Just get a regular vm ?

Sadly I think the correlation between this benchmark and performance is starting to break down imo. Still a legendary idea will be remembered and ingrained in the models forever haha

Honestly happy he’s back at a foundation lab. He will have insane impact there. Of course one of the best educators in the world it’s a bit sad he gave up building and education tool.

Ai in my mind is a new primitive of computing, like compute a db blob storage.

There is a great recent episode of latent space about a similar topic it’s worth a watch even with the click baiti thumbnail and title https://youtu.be/9d899Ram9Bs?is=pQMoVmlWVsTNKfRK

What’s the default context window? Seems extremely short.

Main reason to me is that its layers on layer on top of the base LLM calls with not so much to show for it. Also a lot of native features (like for examples geminis native structured responses) aren't well supported.

Because we are incapable of measuring developer productivity.

Im honestly not that much worried there are some obvious problems (exfiltrate data labeled as sensitive, take actions that are costly, delete/change sensitive resources) if you have a properly compliant infrastructure all these actions need confirmations logging etc. for humans this seemed more like a neusance but now it seems essential. And all these systems are actually much much easier to setup.

Please I need 3 in ga…

Local really has a strange meaning when most of what these things do is interact with the internet in an unrestricted way

It’s just fake accounts being created the bot network has a bot problem haha

There is son much personal info in here it’s wild.

Hmm so that post also links back to segmentation done by structured outputs? (Though here not even enforcing the structure)

https://ai.google.dev/gemini-api/docs/image-understanding#se...

Oh I didn’t know about the visual bounding boxes this is super cool!

Quick question are you talking about this feature?

https://docs.cloud.google.com/vertex-ai/generative-ai/docs/b...

Because it’s just using structured response so it should be doable with Gemini 3 ? (We are using Gemini 3 for some docs processing and its visual understanding is just incredible)

It’s funny how every podcaster/public ai figure is so certain text as a Ui will go away and it’s not going anywhere.

3 flash is also insanely good even slightly outperforms 3 pro for me.

Super excited about this generally ok satisfied with pyright but so I was with conda before uv or black before ruff.

When I was debugging through f# code they definitely had that.

I see this, a hot take form my side as someone who is bought in to GCP i quite like being able to put everything on the same billing account / handle it easily through service accounts.

2.5 pro is already excellent at this.

Now compare on free cash flow

Yep fully agreed the main thing is to break apart the systems so any retries don’t lead to issues like you mentioned.

I do still think there is sufficient amount of boilerplate to potentially justify some engine like this.

Cursor has this too

Yes very good point this to me is one of the most magical elements of this loss how it suddenly makes the model "collapse" on one output and the predictions become sharp.

Yes you are right it's applied to the parameters, but other models (like ngcm) applied it to the inputs. IMO it shouldn't make a huge difference main point is you max differences between models.

To encourage diversity between the different members in an ensemble. I think people are doing very similar things for MOE networks but im not that deep into that topic.

The thing is that regular weather forecasts are also not that great.

Im pretty deep into this topic and what might be interesting to an outsider is that the leading models like neuralgcm/weathernext 1 before as well as this model now are all trained with a "crps" objective which I haven't seen at all outside of ml weather prediction.

Essentially you add random noise to the inputs and train by minimizing the regular loss (like l1) and at the same time maximizing the difference between 2 members with different random noise initialisations. I wonder if this will be applied to more traditional genai at some point.

I wrote som truly awful code back in the day because of that but god it was glorious.