PKE

@thsottiaux @KarolCodes Make a real Codex orchestrator chat that can control, create, read other chats. One that looks at github and delegates by itself so you can truly be the head, but you also have a leader working under you.

@KarolCodes Ship it

Imagine the world where we get this model picker

@Ksparth12 Usually work on 6-8 worktrees

@EllefsenPK tf , how many codespaces u were working in ?!

Didn't realize I'd use Codex this much. 6.76B tokens in 32 days, with a 589M token peak. How's your usage?

@mark_k Yes. Used my Codex x20 up in 2,5 days. Also used both resets so have been this way for a few weeks

I'm hearing it everywhere: The Codex usage limits have had a shrinkflation. It's being discussed all over X. But so far no official confirmation from OpenAI one way or another! Have you personally noticed reduced limits in Codex ??

@thsottiaux Also - Let me structure my archives.

What should we improve in the Codex app. What's not delightful?

@thsottiaux Let me restructure chats. I need to be able to move them from one project to another.

@realBigBrainAI @grok Ilya Sutskever vs Andrej Karpathy

Elon Musk personally recruited Andrej Karpathy from OpenAI, telling engineer Jim Keller he was "arguably the number two guy in the world in computer vision after Ilya Sutskever."

@TrappedPredator @dubcreationX @TonyMunozDev @DevLand95213410 100% agree if i had the time. But right now i have too many ideas and too little time. But of course, if you have the time to learn something properly yourself, or just like doing something, then by all means, i respect that approach as well.

@EllefsenPK @dubcreationX @TonyMunozDev @DevLand95213410 Sure. The line between using technology to help you make art and giving over agency to the machine is thin. Some of it is subjective for sure. Over all I don't really want to get there faster using that technology. Also figuring things out is just too much fun! 😃

Is anyone building games without using AI at all? No code help, no design help, no debugging help, nothing? Genuinely curious.🤔

@TrappedPredator @dubcreationX @TonyMunozDev @DevLand95213410 Codex doesn't have to make any creative choices. It can assist you making your own creative choices come to life faster.

@dubcreationX @TonyMunozDev @DevLand95213410 Might have been a bit poorly worded. Of course using a computer is fine. I just don't think you should leave creative choices to a machine. There's way more fun in doing it yourself. Plus of course there's the effect on the environment.

@flowersslop But it to sleep and use Wake-on-LAN app on your phone

Codex should be able to start my computer when Im not home and shut it down again when its done, in case I quickly need something. I dont know how this is achievable, but it would be really cool.

@BeanJuiceStudio Being good at identifying what makes your game bad and what would actually make it better

Game dev question of the day What's the most important skill for making games?

The genre that has the most potential is friend slop. Despite the explosion of games in the genre there's a severe lack of multiplayer games that are both good and chill. There are few games that are actually like REPO and LC, and both of those are miles away from their potential.

What unpopular game genre is going to have a come back? Go!!!

@vivoplt Playtest to see the result of the code in a 3d game and write down what needs further tuning

TELL ME ONE THING YOU CAN DO THAT CLAUDE CANNOT

@sama All i want is a better, more efficient memory. Everything else is secondary.

i keep thinking i want the models to be cheaper/faster more than i want them to be smarter but it seems that just being smarter is still the most important thing

@rezoundous The newest model perhaps

Soon, not everyone will be able to afford Codex or Claude

@karpathy @karpathy does the future scare you, or do you try to stay in the presence and enjoy the ride for now?

With the coming tsunami of demand for tokens, there are significant opportunities to orchestrate the underlying memory+compute *just right* for LLMs. The fundamental and non-obvious constraint is that due to the chip fabrication process, you get two completely distinct pools of memory (of different physical implementations too): 1) on-chip SRAM that is immediately next to the compute units that is incredibly fast but of very of low capacity, and 2) off-chip DRAM which has extremely high capacity, but the contents of which you can only suck through a long straw. On top of this, there are many details of the architecture (e.g. systolic arrays), numerics, etc. The design of the optimal physical substrate and then the orchestration of memory+compute across the top volume workflows of LLMs (inference prefill/decode, training/finetuning, etc.) with the best throughput/latency/$ is probably today's most interesting intellectual puzzle with the highest rewards (\cite 4.6T of NVDA). All of it to get many tokens, fast and cheap. Arguably, the workflow that may matter the most (inference decode *and* over long token contexts in tight agentic loops) is the one hardest to achieve simultaneously by the ~both camps of what exists today (HBM-first NVIDIA adjacent and SRAM-first Cerebras adjacent). Anyway the MatX team is A++ grade so it's my pleasure to have a small involvement and congratulations on the raise!

@Eoin_Kirwan_ @heydave7 Probably highest chance of success with football player (soccer)

@EllefsenPK @heydave7 Actor less so....

In the age of AI, I wonder if the role of parents in their kids’ career choices is changing. Prior to AI, the traditional view of parenting was to help your child graduate high school and go to college (or at least get into a trade). But with AI, things are shifting. AI is taking over various areas of knowledge work, so just getting your kids “educated” doesn’t seem to guarantee a good job long-term. I think it can be helpful to divide jobs into a few categories: 1. Knowledge work: Jobs where your intellect makes you money. 2. Physical labor: Jobs that predominantly require manual work. 3. Capital: Letting your capital make you money, whether via real estate, interest, equities, etc. 4. Capital management: Probably a spinoff of knowledge work, but focused on managing assets. 5. Risk-taking and people management: Are these higher-level skills than intellect? Maybe. But there is definitely a case to be made for entrepreneurship, hustle, and good people-managing skills. As we dive deeper into the AI era, parents need to think deeply about what kind of jobs will be available for their kids as they become adults and how to help them prepare. Is the best approach still just getting your kids into college and letting them figure it out (the typical approach)? Or does the changing world require a changing role for parents—perhaps one that is more proactive and engaging, where the parent and child work together to find the right career path? Also with AI, how many times will it disrupt what we know about careers? Perhaps every 5 to 10 years? One idea is to have parents go all-in with helping their child develop super high creativity, amazing people-management skills, strong hustle, and high entrepreneurial skills. Will that be enough to allow their child's career to be enhanced by AI and not disrupted? Another idea is relying on capital—letting the parents' capital provide income to their children. Typically, this has been largely frowned upon by culture. Wealthy families sometimes use workarounds where their adult children manage the family’s real estate or assets, so there is still some work getting done. Some parents also bring their adult children into their business as partners or apprentices, grooming them to take over one day. There’s also the concept of parents actively working alongside their adult children to help them find jobs or adapt to changing industries. This is interesting. Perhaps the parent’s role simply has a longer horizon now than it did in the past. I'm not sure what the right idea is, but I wonder who else is thinking about this. Who else is writing about it? What are the most innovative and interesting ideas out there regarding all of this?

@teslashanghai @garyblack00 Neither. Compared to its earlier progress. 2025 - 2026 made way more progress than the average of 2018-2025

@EllefsenPK @garyblack00 Compared to your expectations or compared to the time Elon made public?

Shorting stocks is no picnic. We short stocks with businesses facing secular demand decline or permanent market share loss and where we perceive the company lacks the tech, brand, distribution, or management depth to turn it around. We won’t short a company just because it looks expensive - instead we just won’t own it. We generally don’t short stocks with more than 10% short interest. Shorting stocks successfully is as much about the timing of the short as the short thesis itself. To put on a short position, we need a specific thesis and catalyst over a defined time period. We wouldn’t short $TSLA even at 198x 2026 Adj EPS. It’s too good a company in a thriving business (globally, EVs are growing at 20-25% per year on their way to 80% adoption from 24% today), and TSLA’s marketing issues are easy to fix (launch a conventional pickup truck, advertise, educate consumers on benefits of FSD). We fervently believe that $TSLA solving for unsupervised autonomy will sell more Teslas.

@HansMokeNiemann Love the new hairstyle! Very bold and fearless. Go get em! 💪

Honored to return to Prague. Really looking forward to playing such a strong event.