Jerry Tworek
@MillionInt
Followers
24K
Following
2K
Media
130
Statuses
2K
Berry farmer @ OpenAI | o3, o1, GPT4, ChatGPT, Codex, Solved Rubik’s cube with robotic hand | cautious AI optimist
San Francisco, CA
Joined January 2013
We trained a model and it is good in some things
We're releasing a preview of OpenAI o1—a new series of AI models designed to spend more time thinking before they respond. These models can reason through complex tasks and solve harder problems than previous models in science, coding, and math.
28
55
1K
Haskell is amazing because you can name functions liftM2 and foldr1 and everyone is chill about it
2
2
36
Why Adaptive Bitrate Streaming (ABR) supercharges AIOZ Stream? Our player adjusts quality in real-time to your device & network speed, delivering maximum QoE (Quality of Experience). Key Features: → Global DePIN devices for minimized latency → Increased redundancy during
11
25
115
That reminds me of some tweet….
A thing often in common among great startup investors, founders, and researchers: Trading making a lot of small mistakes in exchange for getting a few giant wins. (Surprisingly many people seem to prefer a few big mistakes in exchange for a lot of small wins.)
2
1
114
Ablations are for the weak
ablations are for the weak. just yolo your runs. (ok, do some small amount of ablations, but don't over do it). instinct is everything in ML and AI.
2
2
93
yes
when i've observed the greats operating at their absolute best they resemble cracked out starcraft players to me. at the top of their game they embody action alone and forget causes, consequence, circumstance. this looks psychopathic from the outside: how does he not crack under
1
0
43
@karenvaites Honestly, my takeaway from this isn't that tracking is bad, it's that teachers can't be trusted to do it right. Which, functionally, might be the same thing.
0
0
14
If a computer makes a bad management decision, it will be met with heavy gradient update Its more than can be guaranteed about many humans
Computers are beginning to make management decisions. Therefore we must begin to hold computers accountable for their decisions.
14
20
351
It doesn't matter how efficient you are at optimizing the wrong objective
ML Systems people need to be stopped. Half of these kernel fusions are not numerically stable 😭 Yes it makes GPU go brr but it also breaks policy gradient theorem and makes me question my life decisions every day
3
1
55
🤣🤣🤣
researchers when asked to switch from bf16 to fp16 and do loss scaling because it is way better for RL
0
1
45
Advantage of founder-led companies is the same as of original authors of the codebase They know all the context on why certain decisions were made Whoever inherits complex systems always has this dilemma "if I change this weirdly looking thing will all everything collapse"
8
3
105
So, I just got back from Hangzhou, China; Attended @IROS2025. Takeaway: China’s robotics ecosystem is moving faster and more coordinated than the US
14
8
48
Codex is far from perfect but the ceiling is actually very high what it can do with good handling. First real agentic product in the world
codex, make it faster >okay done okay codex benchmark dot py says that its faster. can you make it more faster? >okay done okay cool. its faster again. can you add more timing logs so we can make it even faster >okay
10
8
158
I don’t know what kink this is but it have a hard time saying "Grace-Blackwell superchip” without a shiver of excitement going down my spine
4
2
66
Humans have been sycophantic and reward hacking since forever
2
3
168
In the world of online learning every eval is instantly contaminated Train test split is an anachronism of old days, theres only past and future
6
3
148
I’ve always been a great fan of vertical integration - I’ve found inefficiencies around company boundaries jarring TSMC proves me wrong, that in some cases it is a vastly superior strategy to go for a thin slice of a wide market amortising capex through all of the economy
9
7
175
Real tragedy is most of those are not even coming from good models
This is the greatest flippening of all time. Artificial language models have surpassed biological language models in terms of content generated.
13
6
189
The human mind craves stories and makes them from everything around it
9
5
94
It’s a good agent
GPT-5 Codex is the best launch of Q4 2025 it follows instructions, sticks to the guidelines, keeps things simple, and produces optimized code if you're into vibe coding, it might not be for you. but if you know what you want, it beats claude code in every way
5
4
164
This is the way
I’ve got agents doing some Ghostty work while I’m in an Uber. I’ve got some rote refactoring I’ve been wanting to do and now’s the perfect time. I just review the work every so often.
4
2
64
RL really feels like a technical revolution within revolution. It spun up a completely new wave of startups, products and thought leaders on top of a huge wave we were already riding on
16
7
230
Codex is giving me factorio dopamine hits times a healthy multiplier
managing fleets of agents should be more fun than playing factorio with the UI/UX to boot
12
18
406
Technical decisions matter kids
Who would have thought that a multi trillion dollar cap company could have been thrown into such chaos (layoffs) by a single technical decision they made a year ago - using expert choice MoEs for their frontier model.
3
8
187