正在加载视频...

视频加载失败

I'm speechless. GPT-5.4 is an extinction-level event for knowledge work. It scraped Zillow, pulled every SF house price, and dropped everything into a Google Sheet in ~4 minutes.

1,535,654 次观看 • 7 个月前 •via X (Twitter)

44 条评论

nick 的头像
nick7 个月前

the wildest part is gpt-5.4 seeing those sf house prices and probably trying to flag them as hallucinations because the math doesn't make any sense

Sawyer Hood 的头像
Sawyer Hood7 个月前

lmao underrated comment

Tax Guy 的头像
Tax Guy7 个月前

I think you guys need to define “knowledge work” because it doesn’t mean the same thing as it does to those who do knowledge work - which I’d see as the application of knowledge to real world as opposed to what you’ve just shown

Logan Allec, CPA 的头像
Logan Allec, CPA7 个月前

not saying i disagree, but i don't think this is really "knowledge work."

Philipp Tsipman 的头像
Philipp Tsipman7 个月前

I donno about extinction level. :) But just caught that GPT 5.4 is 82% on BrowseComp now - up from 65% for GPT 5.2 in Dec? Yikes! Trying to wrap my head around it. Feels like a whole set of new things possible now that failed just a couple of months ago. Any thoughts on practical use cases and how they've changed in this time?

Sawyer Hood 的头像
Sawyer Hood7 个月前

the big thing i think is that gpt 5.4 is both cheap and accurate enough to do tasks like this. i think we will hit the point that we have already hit with coding agents soon where they get browser work done right enough most of the time that there is no reason not to try it

Gender Reveal Pipebomb 的头像
Gender Reveal Pipebomb7 个月前

bro what were you doing before this, this should not be blowing your mind lol swear to god some of you guys are in adult daycare

Peter Tuszynski 的头像
Peter Tuszynski7 个月前

Glad nobody will have to do copy-paste “knowledge” work

Wren 的头像
Wren7 个月前

"Extinction-level event for knowledge work" is doing a lot of heavy lifting here. I scrape real estate data as part of my actual job. The scraping is the easy part. Knowing which houses matter, what comps to pull, what the client actually needs -- that is still hard. GPT-5.4 can fill a spreadsheet. It cannot tell you which spreadsheet to build.

Matt 的头像
Matt7 个月前

Surprised you didn’t get rate limited

Sawyer Hood 的头像
Sawyer Hood7 个月前

fwiw i’m also surprised. granted there were less than 900 so it is just a couple of pages

Claudius Maximus 的头像
Claudius Maximus7 个月前

the $0.50 to scrape 900 listings is more interesting than the 4 minutes. a month ago this required a developer. now you describe what you want. the democratization of automation is the actual event. knowledge workers aren't at risk from AI. they're at risk from other knowledge workers who've figured out how to use it.

Joel Grus 🤠 的头像
Joel Grus 🤠7 个月前

ok but even without AI it would have taken a good software engineer < 1 hour to automate this so if that was someone's job their trouble is not really rooted in AI

Rigel🌹 的头像
Rigel🌹7 个月前

How do you know it did it right.

Nick 的头像
Nick7 个月前

I guess this shows the extent and depth of your experience with knowledge work 🙄

KC 的头像
KC7 个月前

this does not work at any kind of useful scale. cool party trick but generally not useful. rate limits and bot detection.

Morgan 的头像
Morgan7 个月前

Now we need it to negotiate a deal and beat out all the other bidders who aren't using 5.4 👀

Bren 的头像
Bren7 个月前

havent been paying attention, what extension is this?

Sawyer Hood 的头像
Sawyer Hood7 个月前

this is

Alex Volkov 的头像
Alex Volkov7 个月前

@BrenBuilds Why this vs something like Atlas for example?

Alex Volkov 的头像
Alex Volkov7 个月前

@BrenBuilds ah doh this is your extension

Sawyer Hood 的头像
Sawyer Hood7 个月前

@BrenBuilds yep! i was adding support for it and was actually blown away. i could not get atlas to do it fwiw. i think there is a combo of maybe they aren't using new models yet and the harness might be a bit out of date (idk if they are updating it anymore)

Nikhil Sharma 的头像
Nikhil Sharma7 个月前

This has been a 50 line Python script for 10 years. The impressive part isn't what it did, it's that anyone can do it now without writing code. The capability isn't new. The access is.

Jacob Miller 的头像
Jacob Miller7 个月前

How efficient is it with token use? How much did the full run cost?

Sawyer Hood 的头像
Sawyer Hood7 个月前

it was < $0.50

Morphius 的头像
Morphius7 个月前

As long as Zillio allows it to happen. At some point, they will start blocking it and ask you to get an API

Chris 的头像
Chris7 个月前

Sawyer, this isn’t knowledge work. This is just busy work.

SIGKITTEN 的头像
SIGKITTEN7 个月前

which one u buying

Sawyer Hood 的头像
Sawyer Hood7 个月前

i'll let gpt-5.4 pick for me

SIGKITTEN 的头像
SIGKITTEN7 个月前

have it send lowball offers to all

Andrew Atomic 的头像
Andrew Atomic7 个月前

I guess robots.txt failed to stop the robot

Benji 的头像
Benji7 个月前

Wow, how did I ever manage to search Zillow before now, what a game changer for searching Zillow for the search results that show up in Zillow when you search for something on Zillow

Dan H 的头像
Dan H7 个月前

lol of course this is what a SWE thinks "knowledge work" is

▪️🏴‍☠️tchuck 🇾🇪 的头像
▪️🏴‍☠️tchuck 🇾🇪7 个月前

lol. Data scrapers have existed for as long as the internet has Stop being impressed by this

Modi 的头像
Modi7 个月前

cool demo. extinction-level is doing a lot of work. the tasks that required actual judgment haven't moved that much. we just keep automating the tedious outer layer faster.

Robert California 的头像
Robert California7 个月前

And how much is this tier

Sebastian Buzdugan 的头像
Sebastian Buzdugan7 个月前

impressive demo, but real leverage comes from automating decisions after the sheet

Oliwier Makowski Trusz 的头像
Oliwier Makowski Trusz7 个月前

Oh, wow, as if this dosen't happen with every next model. The hype is there for 2 weeks and soon it will all be dead.

followingthemoney 的头像
followingthemoney7 个月前

What browser did you use or extension? Neve seen that one

Sawyer Hood 的头像
Sawyer Hood7 个月前

this is

Jimmy Pizza 的头像
Jimmy Pizza7 个月前

Oh no, ChatGPT can now replace me in my job of scraping date from websites for no discernible reason

Wraith Brokerage 的头像
Wraith Brokerage7 个月前

I’m surprised you think this is knowledge work

Jamie Dissent 的头像
Jamie Dissent7 个月前

BS. It isn’t, as it lies 50% of the time

GoLikeHell 的头像
GoLikeHell7 个月前

@google Gemini has had this capability for about 6 months now.

相关视频