Well, if you haven't tackled a problem before, then you learn by trail and error - a bit of curiosity and give it a go.

Now, I'm not Einstein, but it seems that if you're just pre-trained, and can't learn, then learning is going to be a problem.

"Reasoning models" achieve their reasoning gains by using RL training to encourage (i.e. narrow) choices of reasoning steps, and sequencing, to those that lead to correct outcomes (in domains like math and programming where this is well defined and testable).

Should we be surprised that if problem statements are tweaked not to match those seen in training that this biasing towards "good steps for this problem" doesn't work as well ?!

Most of these "reasoning" models (i.e. ones using this approach) don't have any fundamental enhacements to algorithmically do better reasoning - just to pick better reasoning steps for problems of a given type they were trained on.

The US government is now officially for sale.

Is the White House available as an AirBnB yet ?

An LLM providing a summary isn't the same as summarization!

To provide a real/meaningful summary (retaining the key points and discarding irrelevant detail) requires reasoning and understanding of the subject matter and context as well as some understanding of what is important. Humans can do this ...

Howevr, an LLM providing a summary is only PREDICTING what a summary would look like - it will give you something that looks like a summary in terms of form and length, with some references to the source material, etc, but is liable to miss what are the actual key points to be retained.

It's a bit like asking an LLM how the farmer can cross the river in least trips with his chicken and corn/etc, and getting a reply that LOOKs like a solution (first cross with A & B, then return with A & C, etc), but in fact is garbage with redundant crossings with same items back and forth. Sometimes they solve these problems successfully (especially "reasoning models", but other times they get overly hung up on the formulaic type of response expect, and generate based on form vs semantics).

Remember, an LLM is a *prediction* (copy cat) technology, it isn't AI or AGI no matter how many times you call it that.

The purpose of advertising and modelling isn't to show you how YOU would look in those clothes (which AI can now do for you - Google has virtual try-on), but more to be aspirational - if you buy our brand then you too will have this amazing lifestyle. All advertising is the same - not just fashion.

I'm not sure AI models for advertising etc will stick, at least not for major brands. There's a reason companies like to hire celebrities (movie, music, sports stars, etc) for advertising campaigns, and like to use the same small stable of supermodels (i.e. also celebrities).

In the fashion industry I'm pretty sure that building a brand image trumps having a stick figure to hang your clothes on, and part of that is associating your brand with celebrities whose image you want to be associated with. Look at the success of Sydney Sweeney's "great genes" advertising for American Eagle - I don't think they'd get a fraction of the buzz if it was an AI model instead.

Yeah, one can complain about stereotypes, but there are definitely "cultural issues" with Indian developers. Plain speaking and honesty don't seem to be that common, and as you say deflection of responsibility vs taking ownership is also common.

Ever notice how many sick days Indian developers take, or how often they have "laptop issues", etc ?! Of course there are power and internet outages, but a company laptop is a laptop ...

It's funny to hear Indian vendor management tell the team "I don't want anyone reporting sick - we need to get this done". Everyone knows the sick days are a scam, and for the most part just accept it.

> Or perhaps to muddle through and pretend that the commented contracts are irrelevant

I'm guessing that the only way to get graded as correct.

If you took the method definition at face value you'd be entirely justified (given no other comment about error handling) to throw an exception if this assertion is false, and your code would therefore fail if they test it (unless their test code is doesn't even test for their own corner case).

Either these problem statements are deliberately written to be confusing, or whoever wrote them should be fired.

Are they testing for programming ability, or for ability to read poorly worded problem statements ?

I have a hard time believing that writing for loops to access an array is causing problems for students, but rather the excessively verbose description of the clearPair() method, including all the examples which might make one assume there is some corner case under-specified by the description.

It seems like middle school math problems where the math itself is easy but the challenge is understanding the problem statement.

Valuations of AI companies do seem to be in a bit of a bubble - priced on promise/dreams rather than any reasonable profit projections, but I expect the ones to be hurt when the bubble bursts will be the pure AI companies like OpenAI, Anthropic and X.ai.

Microsoft's valuation isn't derived from AI sales projections, nor is that of Google, Amazon or Meta.

Can you imagine discovering that you were conceived 30 years before you were born, and have a twin sister 30 years older than yourself ?!

For that matter, what kind of nutjob people want to "adopt" a 30-year old embryo ...

The real miracle will be if this kid turn out normal.

Perhaps, although evidentially his supporters are also plenty mad about Trump's Epstein cover-up.

I think this has been simmering for a long time - part of the whole Q-anon thing from Trump's first term, which he did nothing to shut down, was that the government was full of pedophiles, and it seems his supporters were expecting the Epstein files, which Trump promised to release, to contain the some of the names ... But of course when Trump realized how prominently he himself featured in the Epstein file (perhaps videos too? - something spooked him), then he decided that protecting his own ass was more important.

Billions in stock buybacks is tenths of one percent of a multi-trillion market cap, so irrelevant.

You obviously don't like MSFT, which is fine (I'm not fond of them either, although I do admire the turnaround job Nadella has done), but you should be realistic and realize that the investors moving a trillion dollar stock are not idiots - they are fund managers backed by departments full of analysts with an infinitely greater understanding of the company's financials than you will ever have ... Like it or not, the company is doing well financially.

Sure - I don't expect to see them do it in the short term, but I doubt they'd have much trouble finding a buyer at that price since that's the valuation that OpenAI are taking new investments at.

I'd hardly characterize AI as MSFT's "last great hope" - their cloud business is going gangbusters, Windows continues to be a cash cow, Bing/Edge/advertising is doing fine, Github/Copilot is doing well, etc. Assuming AI usage volume continues to climb, then I expect the cloud providers serving it will be one of the big winners - as long as they have competitive models to serve.