Sam Abrika (@SamAbrika) profile picture

Entrepreneur | Investor | I write without AI

Joined November 2017
Photo by Sam Abrika (@SamAbrika)Photo by Sam Abrika (@SamAbrika)Photo by Sam Abrika (@SamAbrika)Photo by Sam Abrika (@SamAbrika)Photo by Sam Abrika (@SamAbrika)Photo by Sam Abrika (@SamAbrika)Photo by Sam Abrika (@SamAbrika)Photo by Sam Abrika (@SamAbrika)Photo by Sam Abrika (@SamAbrika)Photo by Sam Abrika (@SamAbrika)
AI labs have invented a brilliant upsell: pay more and your AI will think harder. I tested different reasoning effort settings on models trying to navigate financial traps and contractual fine print.
1
8
Most models were already getting almost everything right at low effort. Giving them more thinking time barely changed the results. GPT-6 Luna was the exception.
1
8
So we've arrived at a pricing model where the customer pays for how hard the AI tries, rather than whether it gets the answer right.
5
Barclays is cutting 5,000 jobs “because of AI”. Depending on who you listen to, this is either the start of mass unemployment or a small step towards Elon’s world where nobody needs a job because money will grow on GPUs.
1
21
Most of the cuts are in support functions and IT. Ironically where Barclays was the worst… When I was their customer, logging into my account was always a pain: authentication breaks, app glitches, and customer support can’t fix either.
1
33
Broken technology was the top creator of jobs over the past 30y. So the AI apocalypse is LLMs replacing the humans we’ve been using as middleware between pieces of software that were never properly tested.
10
Sam Abrika retweeted
Capitalism vs. Socialism
197
1,170
80
7,528
157,692
I think I’m becoming a wicked fintech bro. I can’t read annual reports without looking for ways to trick my AI agents. Not to make money (obviously), just to settle the deeply important question of who is the smartest.
1
7
I was expecting a massacre. Well… no. Six of them got almost everything right. Even the cheap models, which I found frankly disrespectful.
1
2
No gotcha moment :( Instead, I discovered the limits of my own ingenuity, which is a less flattering finding and probably the more useful one.
4
LLMs get the facts right, just not all of them.
1
24
In 7 years of JPMorgan calls, Haiku counted 9 quarters mentioning COVID. Except there were only 15 in reality. Relief for hyperscalers, the world still needs large LLMs… for now.
1
6
So AI is getting smarter, but its remaining mistakes are harder to spot. You can check a number against the filing, but good luck spotting a missing fact if you didn't already know it. Too early to retire, humans are still needed.
3