laborcontract

You’re fundamentally misunderstanding the difference between biology and physics.

Medgemma gets its domain expertise from pre-training on medical datasets, not finetuning. It’s pretty uncharitable to call the post an awful take if you’re going to get that wrong.

You do you, and if it works i’m not going to argue with your results, but for others, finetuning is the wrong tool for knowledge injection over a well-designed RAG pipeline.

Finetuning is good for, like you said, doing things a particular way but that’s not the same thing as being good at knowledge injection and shouldn’t considered as such.

It’s also much easier to prevent a RAG pipeline from generating hallucinated responses. You cannot finetune that out of a model.

This is simply gatekeeping. I understand writing an article about it but filing a radar?

The only thing the author wants to do here is a fandom flex.

Ironically, his pictures prove that the icons have changed a lot! What’s the canonical face anyway? You can look at the prior icon and argue they broke canon there removing the black center stroke and clipping the overflowing line.

I totally understand the desire to hate the liquid glass aesthetic but at least the downsides are mitigable by disabling transparency.

On the flip side, having used the beta for a while, I think I’ve fallen in love with it. The UI is absolutely gorgeous and such a delight to use. It’s the closest thing we’ve come to Aqua’s “lickability”. It’s not something that can be judged by product videos and screenshots.

From a purely utilitarian perspective, there are some regressions. The most egregious one being the lack of a properly sized url bar in Safari.. but on the other side of that, content and websites are so much more immersive with this design.

We’re definitely going to see improvements made over the beta period. Notifications are a swing and a miss and impossible to read.

I really thought I’d hate this new design language but I think Apple’s on to something here. I remember the good old days of the first iphone, and how it engaged you in a way that made you want to do stuff with the UI like flick through cover art, scroll up and down my music library just for fun, and pinch to zoom in on random things.

The liquid glass redesign is recapturing my imagination in the same way. I’m fiddling just to fiddle. This is not vista.

I don't think so. It could possibly hint at a reduction in cadence for releases. There has been an open line of reporting about Apple that yearly releases were important to Apple but that the yearly release schedule also made it hard for them to divert engineering resources into more experimental projects like LLM-related development.

Linking version numbers to years could be seen in effect to make each one seem less monumentally important.

Link to underlying link rather than to the quote-tweet spam:

https://xcancel.com/ruxandrateslo/status/1924206417000403328

  > @RuxandraTeslo: “I hadn't realized this but Romania's next president was 1st in the world in the International Maths Olympiad 2 years in a row with maximum score”

His name is Nicușor Dan. He was first in 1987 and 1988. https://wikipedia.org/wiki/Nicu%C8%99or_Dan

is there a way to do this in vscode, btw?

Yep i’ve been using the old and new models in cline. I can’t tell any difference outside of the improvement with diffs, but that’s good enough for me.

Update 2: I've been using this model in both aider and cline and I've haven't gotten a diff matching error yet, even with some pretty difficult substitutions across different places in multiple files. The overall feel of this model is nice.

I don't have a formal benchmark but there's a notable improvement in code generation due to this alone.

I've had gemini chug away on plans that have taken ~1 hour to implement. (~80mln tokens spent) A good portion of that energy was spent fixing mistakes made by cline/aider/roo due to search/replace mistakes. If this model gets anywhere close to 100% on diffs then this is a BFD. I estimate this will translate to a 50-75% productivity boost on long context coding tasks. I hope the initial results i'm seeing hold up!

I'm surprised by the reaction in the rest of the thread. A lot unproductive complaining, a lot of off topic stuff, nothing talking about the model itself.

Any thoughts from anyone else using the updated model?

Knowing when to call functions, generating the proper function calling text structure, properly executing functions in sequence, knowing when it's completed its objective, and doing that over an extended context window.

My guess is that they've done a lot of tuning to improve diff based code editing. Gemini 2.5 is fantastic at agentic work, but it still is pretty rough around the edges in terms of generating perfectly matching diffs to edit code. It's probably one of the very few issues with the model. Luckily, aider tracks this.

They measure the old gemini 2.5 generating proper diffs 92% of the time. I bet this goes up to ~95-98% https://aider.chat/docs/leaderboards/

Question for the google peeps who monitor these threads: Is gemini-2.5-pro-exp (free tier) updated as well, or will it go away?

Also, in the blog post, it says:

  > The previous iteration (03-25) now points to the most recent version (05-06), so no action is required to use the improved model, and it continues to be available at the same price.

Does this mean gemini-2.5-pro-preview-03-25 now uses 05-06? Does the same apply to gemini-2.5-pro-exp-03-25?

update: I just tried updating the date in the exp model (gemini-2.5-pro-exp-05-06) and that doesnt work.

uv solves this problem.

  uv venv python3.11

done.

Disagree gently here. Vibe refactoring is so much harder than vibe coding. I understand the temptation of having LLMs refactor code because that's ostensibly what they're great at – "transforming" the code. But LLM's are opinionated enough where they love to arbitrarily drop features or constraints for no particular reason, and they won’t tell you what they dropped along the way.

Proper refactoring with an llm requires full testing coverage and, by the time you do all that, is the refactoring really necessary? I prefer 100% stability. If you're refactoring the because it's poorly structured and unreadable, that’s okay.. LLMs can help understand it.

In my use of llms, i find it’s actually much easier to rebuild something from scratch rather than refactoring flawed code. It’s much less likely to inherit strange assumptions and code smells that way.

With all that said, the one prompt I do use when refactoring is to tell the llm to do a lossless refactor and then follow up with a "was this really lossless"? It's not foolproof. LLM's love to lie.

Google Cloud Console's billing console for Vertex is so poor. I'm trying to figure out how much i spent on which models and I still cannot for the life of me figure it out. I'm assuming the only way to do it is to use the gemini billing assistant chatbot, but that requires me to turn on another api permission.

I still don't understand the distinction between Gemini and Vertex AI apis. It's like Logan K heard the criticisms about the API and helped push to split Gemini from the broader Google API ecosystem but it's only created more confusion, for me at least.

> This is continued whataboutism

i think i’ve written enough to show that’s that last thing im trying to do.

Obama dropped a bomb on a us citizen without giving him due process. How should this case be handled? Like i said, american presidents causing untold misery on some people is a tale as old as time. The point is to create a system that is, on balance, just.

would you prefer what’s happening in brazil?

there are cases where the president has made the lives of individuals miserable in every presidency we’ve lived through. Stating that this is a terrible fact of life and doesn’t justify the harms.

what kind of powers do you envision the judiciary having to rectify these wrongs?

I’m hoping to shine some light on bad executive behavior, not to create a contrast that makes Trump look good relative to what FDR did.

We agree here.

Because we've seen federal courts unafraid to repudiate unilateral executive force.

Why don't you think that, all things considered?

The internment of the Japanese cannot be brushed off. That was a gross violation of power that, if Trump were to do something similar, would tear this country apart. On top of that, one of the side effects of the internment is that it's sowed a deep distrust in the federal government and the Census that continues to harm representation to this day.

They used Census data to round people up. It doesn't get much worse than that.

All the while, I don't think I've ever loved our country as much as now because I think this is a time where our system of checks and balances can come to shine.

I do hope this experience will lead to people re-evaluating their love of FDR. You can like what he stood for, but he was an equal if not (much) greater abuser of the executive office.

Facebook is one of the few large companies that's in control of someone who is able to take large risks and does that by way of R&D. I find this admirable (do not conflate this with "I find mark zuckerberg admirable").

Just look at what Microsoft did with Hololens. They bought it, chafed at the thought of a couple billion dollars of additional capex to make it a viable product and dumped the entire project 20 meters from the finish line. If you're going to do something, do it with conviction.

Whose money is this? The company's money and investors' money. Let them lose it.

Should I be cheering for them to instead to spend that money on streaming rights? Buybacks? C'mon now.

fwiw apple has been extremely generous to be on refunds and, while I don't abuse the system, I do get a refund when I think it's deemed necessary, either due to a lack of satisfaction or accidental purchase.

Chargebacks are a huge pain in the butt to deal with and, as someone who's saw this first hand, chargebackery is correlated bad customership (two words that I just made up) to so I can understand that they'd hate consumers doing that instead of going through what's otherwise a pretty fair system.

Dealing with apple support in payments land has been, on the consumer side, one of the less infuriating things that have come out of what is now their support process. That said, the ux for getting refunds and checking on their status is antiquated, perhaps purposefully.

It’s three taps to make a website appear as a bookmark. Redirecting the user to the app store is as many taps.

Consider why more companies haven’t gone the way of PWAs – there are a lot more reasons higher than installation ux.

A one year posture shift pdoes not constitute famous hostility. He dismissed app stores because the iphone didn’t have that feature yet. apple was the king of digital storefronts with the itunes store.

OpenAI has the business development side entirely fleshed out and that’s not nothing. They’ve done a lot of turns tuning models for things their customers use.

i've been doing this with my child and picture books. i take pictures of pages, convert it to a coloring book, print it, and then we color her favorite books together.

Yahoo throwing its name into the conversation throws me right back to the bad old days of web browsers, when Yahoo’s hideous toolbar was a staple of the ‘helping your grandparents with their computer and wondering what the hell all these things are doing on the screen’ era.

It’s also making me wonder if we know what we’re really getting into with the chrome divestiture. A Yahoo chrome would make the web a worse place.

…i can’t believe i’ve been running a script to ingest the data for the last six hours. thank you.