Cameron Williams
127.8K posts
Cameron Williams
@Wasgo
AI Critic. RPG Designer. Comic Creator. Fashion Designer. Self-assigned titles are the best. RTs of my own Tweets are endorsements. (He/Him)
Vancouver, Canada شامل ہوئے Şubat 2009
266 فالونگ912 فالوورز
Wear your underwear on the outside and eat raw chicken if it saves you a few minutes.
English
Apparently ChatGPT thinks you should leave the house naked if it saves five minutes.
English
@CentroLeaks Do have a link to the document? These are awfully round numbers.
English
Development Budgets:
- SV DLC: 500 million yen
- Legends 2/ZA: 2 billion yen
- Winds/Waves: 3 billion yen
English
Cameron Williams ری ٹویٹ کیا
@TransLink Up until a month or two ago, I almost always got reliable real-time data on the 33. Now it's 50/50 at best. Is there a reason?
English
@Wasgo Hi Cameron, unfortunately real-time data is not always available on the Next Bus website! You can always give us a call at 604-953-3333 to speak to an agent. They will try their very best to assist you, using the information and resources available to them. ^hb
English
@TransLink How come real-time data is so inconsistent lately? I was waiting for a “scheduled” stop and no bus came. With real-time ones, I don’t have to waste time like that.
English
@NewTSage I need someone to take that and then claim Gal Gadot is the best actress of all-time.
English
I love that numbers = law of the land for you people.
fangirl19 (Rye)@fangirl1759
@NewTSage Yet, Red Notice is the most watched Netflix movie of all time. Rock brings in those numbers regardless, streams or box office.
English
@jpooch21 Newsom is great at beating Republicans but he’s not a great leader. He’s been anti-worker, anti-homeless and anti-trans for too long already.
If he’s the only electable option, sure. But he’s not great.
English
I feel like people are coming around on Newsom, because he’s one of the only elected Dems showing any sort of backbone and fight against the bullshit we’re living through right now. I get the left left isn’t a big fan of his, but I hope they appreciate this.
Brian Tyler Cohen@briantylercohen
.@GavinNewsom: "I think he should start with Shreveport, Louisiana, and Speaker Johnson's district that has 6x+ the per capita murder rate of Nancy Pelosi's San Francisco. What about the carnage, Mr. President, in Shreveport? Why aren't you protecting the folks there? Eight of the top ten murder states in this country are red states. What bullshit."
English
@vig39 @GaryMarcus Weird. I don't know why. Here's a screenshot. I never post these unless I can replicate. Using a name has the replication rate for this one at about 33%. But if you do the question "B's in blueberry?" on GPT-5 non-thinking, you'll get about a 60-70% replication rate.
English
@vig39 @GaryMarcus That query I did more testing on after. But here’s the exact same error on the same query. chatgpt.com/share/e/689c14…
English
@GaryMarcus That’s how you know he’s a troll… none of the reasoner models will make this error
English
@CtrlAltDwayne @GaryMarcus For me it helps me work out what’s a genuine change in how GPT functions and what consists of a problem specific patch.
English
@Wasgo @GaryMarcus absolutely. i think this actually highlights the importance of asking proper questions. this wouldn't be considered a fully formed question in the real world. maybe it's not an intelligence problem we have in the models, but the users instead.
English
@Angaisb_ @GaryMarcus I found this one because I was trying to determine if counting was a model fix or a new tool the thinking model uses. Pretty sure it’s the latter.
English
@GaryMarcus Gary, I honestly was starting to like you again, I even followed you, but making fun of a non-reasoner having tokenization issues that we all know they have... come on
I agree it should've been routed to Thinking, but still, they're improving the router
English
@CtrlAltDwayne @GaryMarcus I’m pretty sure that there’s a new tool the model uses for counting, and if you use the full question of “How many R’s in…” it will always trigger the thinking model.
That being said, yours is hilarious. I love it.
English
@GaryMarcus to be fair though. if you regenerate the response using the thinking model it gets it. but auto mode wasn't smart enough to switch to thinking mode for this question in the first place.
English
Prize for the best meme expressing this increasingly common sentiment :)
Lizzie Sinabox@sinaboxlizzie
@GaryMarcus was right.
English
Just asked GPT-5 to count letters in various words (b's in blueberry, y's in syzygy, r's in strawberry, i's in antidisestablishmentarianism, etc.) and it got 10/10 right. I have no doubt one can elicit poor performance in this kind of question but it's atypical.
English
@josef21m @levelsio This is actually my favourite example. GPT-5 is an incredible tool and the low to mid level answers gets better every model. But it doesn’t “think” or “reason”. It just maps data.
x.com/Wasgo/status/1…
Cameron Williams@Wasgo
@GaryMarcus Found my favourite Chat GPT-5 Thinking fail so far. It’s hilarious.
English
I hear this from my founder friends a lot too
I don't care what all the developer X accounts here say "noooo AI won't do anything to the SWE job market"
It's 100% coping, because it already did!
It wiped out the low to mid-tier of the SWE job market
You don't hire the low to mid-tier SWE because your existing engineers can do the same job by telling AI to do it
You do hire the 10x engineers for the more difficult and advanced problems that AI can't handle (yet)
Gordon Leete@GordonLeete
The bar is incredibly high. Folks only want a 10x engineer. Anything else - well - there’s AI for that. Senior eng is the new Staff. Even an A tier l33t coder - systems design interview, live debug, and optimization/edge case must be perfect. There is a 10x expectation vs 5 years ago.
English
@littmath I somewhat agree. I think generative AI is an incredible tool. 10 years ago nobody would think we’d have a tool that could write, solve problems and do research.
Conversely though, Altman continues to falsely treat it like it has internal reasoning or human-like intelligence.
English
FWIW I get the urge to post “AIs do the darnedest things” screenshots, but I think the primary purpose they now serve is to fill a demand for reassurance that the models are much less capable than they actually are. Better to try to accurately convey capabilities.
English
@littmath Reading through the "thinking" notes, I would presume they built an internal tool for correctly counting now. I wouldn't expect to be able to find consistent errors in this area anymore.
English