alansaber

"Don’t ask it to show its reasoning in the reply" “Explain why you chose this approach in three sentences” says it all really

Agreed as well, this is my preferred framing, it only makes sense to get into semantic vectors once you realise the very real limits of discrete vectors and bag of words.

Just sounds like a more user friendly interface for projects (rather than a folder, how about the adorable green dot for my cybersecurity questions?)

AI companies rushing an implementation? Surely not :).

I like video as a format because it's pattern-breaking. Consuming thousands of very similarly produced videos sounds even more depressing than consuming LLM blogs and articles. EDIT: That aside, this is of course extremely impressive.

When they inevitably drop allocation after post-launch hype dies down.

Always key to include the one bench where the smaller model inexplicably outperforms the larger model

They're building for future capabilities. It's a logical approach to leveraging the data they have.

I don't much care for it on desktop either

Given how LLMs access compressed knowledge from their model weights, the similarities make sense

Exactly, competition between both frontier model companies and the chinese labs is the primary factor suppressing consumer prices.

Outdated/bad CICD is surely extremely common in every org

So I generally agree with this with some exceptions: these proficiencies will still exist, but move away from standard engineering roles, to highly specialised individual roles, just as there are still specialists in assembly and other low level languages. They'll just have roles maintaining the nuclear arsenal etc etc where artisanal skill is still an absolute requirement.

"a lot of the models work was essentially coordinating everything." - I don't see anything wrong with that personally. It's still extremely challenging to build a model harness, and having a model-mediated everything is clearly wishful thinking. It's exhausting to keep up with, but also somewhat exciting, all depends on your perspective of course.

TBF OP makes a good point: skipping out on the whole fine-tune sing and dance and diving directly into a good general classifier is a barrier to developing actual, deep intuition for the problem space (a pretty prevalant anti-AI argument).

Not even openai want to get flamed on show hn

Recent models have been a big step up on graphics generation. Not exactly sure why, it would be neat if they would release any quantity of technical blogs.

I'm amazed they didn't test xhigh thinking mode explicitly to ensure it didn't exceed the 128k thinking budget allocation. I guess pace of development gets away from everyone, even OpenAI.

I felt that was about 5.5. IMO 5.6 Sol was overindexed: more verbose, prone to overengineering.

It's extremely variable because the products are roughly equivelant, and a lot of the quality of service depends on their inference capacity at any given hour/day.

It would be extremely funny if the explosion in SVG generation capability in particular was a result of this benchmark

Scope explosion. AI is really bad at polish, requiring human intervention. So everyone reallocates effort to breadth not depth, effectively mirroring AI capability.

"New technology: cool but the supply chain is vulnerable". A tale as old as time.

"My project isn't open sourced yet" underlines the problem. There has been a massive explosion of software that works, but badly. The average user experience suffers as a result.

Not dumb questions. These design decisions are based on years of applied testing, more than any theoretical result.

As anthropic/openai subscription allocations get squeezed you'll see more people using "second rate" closed models like grok. The token allowance with a Cursor subscription is crazy.

This is why I like books. No attempt at attention hijacking. Wheras every other modern app tries to claw your eyes out.

Yes but on the sliding scale of how deterministic is this task- writing, not very. coding, can execute and write some tests (even if gamed/bad).

This is a neat workflow. Incentivises you to proactively think about your writing, even as you cut it.

Well said, the issue being, the form factor of many AI products disincentivises this behaviour. To give an example, using a CLI tool with a frontier model to write a blog (actually fairly common), you have to re-open it in another editor, potentially reformat it or convert MD>DOCX, etc etc. There's a lot of noise and mess, and people naturally tend towards "not bother" and 1-shot some "amazing" prompt (because the value of their elbow grease is disproportionately low, unless they actually have something to say- many do not).