Matthew Berman
@MatthewBerman
Followers
71K
Following
20K
Media
2K
Statuses
11K
Building Forward Future. YouTuber, Angel Investor, Developer, AI Enthusiast. https://t.co/9rk7dmIboR
California
Joined June 2007
AGI is a memory problem.
Agent memory is a compression and caching problem.
137
29
698
Superintelligence = . Intelligence ✅ (done) + .Memory 🛠️ (in progress) +.Recursive self-improvement ❌ (not done).
88
24
353
Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.
525
977
5K
What if AI models never truly ‘self-improve’? . Maybe super intelligence isn’t in the models at all, but in the evolving machine-human system: the researchers, software, and compute that keeps building better ones.
118
9
238
The DeepSeek effect on $NVDA was basically free money. It's like the entire world forgot about Jevon's Paradox.
23
1
98
Vibe coding a full Excel clone step-by-step with GPT-5 is kinda nuts. (I know I'm still far from a full clone, but this is after about 45 minutes of work)
33
8
175
Thinking about buying a billboard in SF and using a shitty comment from my YouTube video as the copy.
5
0
14
YouTube video was sponsored by Abacus AI (they didn't pay me to post it to Twitter).
1
1
8
GPT-5 may be the most polarizing model launch in history. Here are the industry reactions:
10
4
90
Thinking about buying a billboard in SF and using a shitty comment from my YouTube video as the copy.
24
1
83
Legacy models are back. If you loved GPT4o, go use it now!.
2
0
24
So many people in my comments think I'm paid by OpenAI. They are 100% confident. I've never been paid by OpenAI. Being on the "media" side now, I see how many false claims and lies are coming from not only the media, but comments/replies as well. Don't trust anyone. Obvious,.
73
14
336
Why would he be scared? . The limiting factor for most people who don't have laundry machines is square footage, right? . Seems like he should buy robots and offer a premium "wash for you" service.
This guy runs a $100M laundromat business and he’s scared. He should be. The only limiter now is pretraining data. Once we feed more data into Helix, the robot will handle more tasks. The hardware needs no changes - it’s already built to do most tasks humans can.
9
1
38
This is the most san francisco thing i've ever seen and I love every second of it.
7
1
62
Someone should make a cabinet that doubles as a dishwasher so I don't have to move dishes back and forth.
49
3
140
Should OpenAI retire the old models?. Is the simpler user interface worth the tradeoff of less customization?.
FROM THE LIVE SHOW: @mreflow and @MatthewBerman on retiring past models. > Probably a smart move to go to a single model.> This isn’t for power users, but for general users.> Paying for GPT gets complicated when there are so many options.> Most people don’t want to think about
21
0
37
Was OpenAI releasing a frontier open source model the same week as GPT-5 the greatest 4D chess move?.
FROM THE LIVE SHOW: 4D chess or conspiracy theory?. Why would OpenAI drop a cutting-edge open-source model… then launch GPT-5 right after?. > Out of the goodness of their hearts to uphold the open mission?.> Or a genius play by Altman & team?. The pitch writes itself:.> Want the
5
4
27
Fun fact: Did you know 36% of all AI researchers are named Sebastian?.
20
0
86
We might be "post eval" as @theo said. Getting 1-2% on math evals doesn't matter anymore. We need new benchmarks that test the qualities of a model that are more difficult to describe. aka the vibe.
FROM THE LIVE SHOW: on knowing if new models are better. > You have to do the vibe check.> We’re at the point where you just need to use the model — they’re all getting gold on the IMO. > Use it, explore its boundaries.> Figuring out the right workflow between models is key
27
4
114