@sarker's banner p

sarker

ketman hetman

0 followers   follows 0 users  
joined 2022 September 05 16:50:08 UTC

				

User ID: 636

sarker

ketman hetman

0 followers   follows 0 users   joined 2022 September 05 16:50:08 UTC

					

No bio...


					

User ID: 636

This is a fully general argument against any new APIs of any kind - after all, existing APIs are already in the training set.

Nevertheless, LLMs can learn to use APIs they haven't seen in pretraining.

It's not clear to me that they could.

Friends I know at apple, even on important projects, find it difficult to get promoted or get a raise. The corporate culture is extremely focused on siloing people so that they don't find out too much about what's going on.

Simply throwing cash at people in the hopes of getting a good model out of it leads to the Facebook path. Not a single incumbent tech company has produced a frontier model except Google, but really it came from DeepMind which was an acquisition and retains a somewhat separate culture from the rest of the company.

No reason people couldn't just ingest language documentation into the context, or even fine tune an existing base model for their language.

One tech company has somewhat called bullshit. They sit in Cupertino. Maybe they’ve since changed but my understanding is they still call bullshit.

It's less that they've called bullshit and more that they don't have the chops to build a good model themselves so they will simply license one.

But the LLM companies are constantly increasing their inference costs as FLOPs go down in price.

Labs need to cover their training costs at inference time. This isn't a problem if there are no training costs.

On top of that, this isn't really true. For example, Gemini 3 flash outperforms 2.5 pro and costs ~2.5x-3x less per token.

More broadly, the cost to train a GPT-2 level model is 600x lower than it used to be. Algorithmic progress has made massive strides, and that applies to the inference side too.

At least according to Ed Zitron's analysis. Maybe you just don't believe his numbers.

I don't.

If a lab goes pop and has to sell off its assets, training costs are not a problem. Inference costs can be covered with a reasonably priced subscription. If we're stuck with current SOTA models for the next 50 years, software dev will still be changed forever.

Innocents Abroad by Mark Twain.

Gemini 3 wouldn't even generate syntactically valid Java 100% of the time.

Why does this matter? You are running the agent in a loop where it can compile the code and run tests, right?

My wife uses essential oil of vetiver for medicinal purposes

What medicinal purpose does essential oil of vetiver have?

They still make syntax errors from time to time such that the code won't compile

This is basically a non-issue in my experience. The code compiles 99 times out of 100, and if it doesn't, I don't care because I'm not reviewing it until the code is compiling and tests are passing.

You have to carefully review all AI-generated code for mistakes (which negates the time savings), or you will get buggy code sooner or later.

This is true for human written code as well.

I'm not a breathless AI booster - I often find the models taking shortcuts that I wouldn't expect from, say, a conscientious coworker. But the pace at which I can prototype and experiment has absolutely taken off in the past month. Ideas that I've had on the back burner for months but I never got around to trying can be farmed out to an agent who doesn't get frustrated or bored.

The radical policy of putting criminals in jail without segregating the bus would have permitted Montgomery to have avoided the bus boycott entirely.

Meanwhile in the US Rosa Parks made public transit a last resort option for those too poor to care about being stabbed.

Do you believe that the thing keeping people from being stabbed on public transit was that blacks had to give up their seats to whites when the bus was full?

I think this notion of letting people who aren't your children run your foundations is quasi-cuckoldry,

In the long run it's all cuckoldry we are all cucked. The moral compass of the median American today would be totally alien to the median American of 1500, and things don't change much if you look at direct descendants of those Americans from 1500.

Anecdotally, one (broadly) lefty woman I knew didn't believe me when I told her that Americans in fact do not have a constitutional right to obstruct government agents ("so what, they're just supposed to protest on the side? Are you sure?"). I think some people just don't really understand what's allowed and what isn't for reasons that I can only speculate about. We may find this Karen someday shocked to be clapped in irons for attempting to burn down the ICE fulfillment center.

I don't even think you'd have to accept that there's any kind of organic issue underlying the trend; even if it's just overdiagnosis/ overmedication for kids' screen-induced behavior problems, any parent who hauls their IPad toddler to the pediatrician for a developmental disorder diagnosis is a parent who is having a much tougher time with that kid than they anticipated.

I don't think you're cynical enough. One reason there might be overdiagnosis is because having autism or an alphabet disease is highly adaptive - the student gets accommodations, the parents get to have "neurospicy" children, etc. There doesn't actually have to be any underlying behavioral problem at all.

Are you saying that these guys can't just turn the lights green?

It's a little grim that even senior government officials live right on a six lane freeway. But at least ВИНЛаБ is easily accessible.

Which countries have a cliff edge drop post pandemic? Sweden does not.

I expect they'd get their lunch eaten by the firm hiring 99th to 99.9th percentile talent.

The "before" is probably a little tight since you're at six hours of sleep assuming it takes an hour to get dinner. You might have to cut down to an hour or two of vidya every day. Otherwise accurate.

No. But if you're hiring the wrong people, I don't think you will become a successful quant firm.

If hiring the right people has little relation to making money, I don't know why you think they'd care about your GPA.

Whining about the hiring practices of highly profitable and successful firms in a cutthroat industry because they won't give you the time of day is a very bad look.

This is a surprising claim since I've read many studies pointing to reduction of fertility among the parous rather than increased childlessness as the major driver of reduced TFR.

He does not speak English

He does speak English.