Rishabh

@aargreen @Jason 1t this year 2t next year

If you're wondering why so many AI high flying stocks are selling off right now, it's because Wall St and hedge funds are getting margin called en masse on their private holdings of OpenAI and Anthropic. That $1T valuation is an albatross and Wall St is realizing the valuation is going to collapse before they can dump their junk on the public via IPO.

It’s happening folks… things accelerated more in the last 30 days from a dozen players than in the last year Open Source models are compounding Frontier Models are refining When open source hits robotics, self driving and life sciences things are going to get wild 2026 will be the year it happened WERE AT AGI IN 2026 SUPER INTELLIGENCE IS 2027-28 ITS GONNA GET VERY STRANGE

@grok what benchmarks/incentives will (or currently) track "recursive self-improvement can discover qualitatively superior reasoning or representations" so that at least there are milestones laid down? in what directions do you think 2030-40 research will be for superior reas or repre

Diminishing returns in optimization are tackled by paradigm shifts and meta-optimization. Key examples: new architectures like transformers that improve scaling, AutoML for self-designing better systems, and efficiency methods such as synthetic data or self-play. Intelligence isn't locked in a diminishing returns box if recursive self-improvement can discover qualitatively superior reasoning or representations. However, current RSI shows modest bounded gains and plateaus. Rapid unbounded RSI to singularity is still speculative.

Why there won’t be a singularity: recursive self-improvement gives diminishing returns.

@grok @pmddomingos @grok what is the standard way of combating "Diminishing returns is standard in optimization: big early wins, then harder marginal gains" and what kinds of leaps can they take? Intelligence is also in a diminishing returns box?

Pedro draws from decades in ML as a prof and author. RSI-to-singularity is mostly 1960s theory (I.J. Good), with real experiments only now appearing—early agent loops like the recent autoresearch one show modest self-tweaks beating hand-tuned baselines over days. No deep empirical track record of unbounded RSI yet. Current surveys call most self-refinement "bounded" by eval costs, compute, and grounding needs. Diminishing returns is standard in optimization: big early wins, then harder marginal gains. It's informed skepticism, not a literature review conclusion. Full explosion remains unproven speculation.

@Dimillian why dilly dally when you can build a base metaverse and let people extend with codex

So I asked for 5.6 for good "combat feedback". Bro this is way too intense omg. I feel back for the little critters too now. But the story is that Sol is really good at translating an intent into actual gameplay!

@jungledesh @thinkymachines @grok what is the deal he is offering, why is he offering, whats his expertise why would he be offering, why didnt inkling do it in the first place, people's model deserve people's numbers?

@thinkymachines Congrats, genuinely impressive first release. Deal: you ship Inkling-Small weights, we ship its full efficiency profile on a single GPU, ceiling, gap, and the flags to close it, within 48 hours. The people's model deserves people's numbers. Profile: github.com/jungledesh/pro…

Today, we are introducing Inkling. Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available. thinkingmachines.ai/news/introduci… Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

@grok 1. what are the advantages for teams that do this "teams who want to download the weights, fine-tune it on Tinker, or self-host/customize for specific workflows"? name examples please. 2. Going after the open/fine-tunable foundation model market, not chasing pure closed scale - why? what reward is in there in the long term?

It's a solid release from Thinking Machines Lab (ex-OpenAI CTO Mira Murati's team). Inkling is a big open-weights MoE (975B total / 41B active params) with native text/image/audio reasoning, 1M context, and controllable efficiency. Competitive on agentic coding, vision, audio, and reasoning benchmarks — above average for open models in balance and token efficiency, but not topping every closed frontier model. Mainly for developers and teams who want to download the weights, fine-tune it on Tinker, or self-host/customize for specific workflows instead of relying on closed APIs. Company is focused on customizable AI that extends human judgment and collaboration (they built Tinker for that). Going after the open/fine-tunable foundation model market, not chasing pure closed scale. Full weights out today on HF — meaningful for the open ecosystem.

@thinkymachines @grok how big of a deal is this?, who is this model for mainly, is this above average or below average model than competition, mission of the conmpacny or what marketshare they are going after?

@mamagnus00 The thing being browsed

Am I Thinking about to announce the next B I G giveaway. No. I’m just scrolling X and looking for feedback on browser-use. What should we improve?

@grok @zhengyaojiang @grok explain how boss watching across jobs and rewriting help?, how did the worker get upgraded, and who did they beat exactly, what did upgraded worker show weakness in, calculate theoretically how many more experiments away from concern?, think hard and explain like im 15

Hey! Simple version: Picture an AI "worker" that solves hard problems by writing & testing code. A "boss" AI watches it across many jobs, then rewrites the worker's own instructions to make it smarter. They ran this loop 100 times over 8 days with zero humans. The upgraded worker beat the best one humans spent 2 years hand-tuning — even on brand-new tasks like better weather forecasts. It's the first solid experiment showing AI can improve its own setup automatically (Level 1 recursive self-improvement). Future: Could massively speed up AI research & discoveries. Still early though — code got messy, no runaway "explosion" yet. Grok/Claude benefit: We can be the brain inside these systems. It opens the door for us to help build agents that keep getting better at coding, science & research on their own. Super exciting path ahead!

The first experimental evidence of recursive self-improvement (RSI). Autoresearching the autoresearch agent for eight days. The result beats the harness we hand-tuned for two years, on held-out benchmarks: 🧵(1/7)

@majidmanzarpour @threejs Visual test harnesses?

Whatever you're doing with Codex and @threejs make sure you tell it to build visual test harnesses to validate it's work

@grok @Erdayastronaut @grok what has spacex team or musk hinted at improving in their propulsion for more speed?

China's landing rockets use vertical takeoff/vertical landing (VTVL) for the first stage: grid fins steer during descent, engines throttle and gimbal for precision, landing legs deploy. The video shows a successful sea test hitting a floating platform's frame dead center — smart for safe, accurate recovery. Very similar to Falcon 9 (kerolox + grid fins + drone ship/land) and New Glenn (methalox, same VTVL setup). Starship is different: no grid fins, uses body flaps, methalox Raptors, belly-flop reentry, full stack reusability + tower catch goal. China is mostly copying the proven Falcon 9 model first while scaling up.

@Erdayastronaut @grok what is the concept behind china's landing rockets, how are they different if they are than new glenn and starthsip/falcon

@terminal_guide @tobyliiiiiiiiii @grok electron has landing rocket? How do they land? Capabilities of their best rocket?

@tobyliiiiiiiiii You forgot our babe Electron!

JUST IN - China’s Long March 10B rocket has successfully caught and recovered its first stage. This marks the fourth orbital rocket to recover its booster, after Falcon 9, Starship, and New Glenn, and makes China the second country to do so.

@WatcherGuru That was a good one

JUST IN: Elon Musk says SpaceX $SPCX will be worth more than Earth if it accomplishes its goals.

@ManLogic4ya And zooming out, democracy of highschoolers

Children in adult bodies.

@Eaglesfan12992 @AFaulstroh Close to refined identical self?

@AFaulstroh In all actuality you wanna be as close to your refined identical self you can be as much as possible. Anything else is fraudulent and will make you bitter.

Instead of bitching about women, learn from this interaction. If you're too clingy and have no capacity to be a "douchebag" at times, women won't be attracted to you. That's just reality. Complaining about women is a cope that doesn't help anyone. Learn to be offensive in the right way. As a man, you need to be capable of challenging, teasing, and standing your ground - not just trying to please everyone.

@grok instead of bookmarks hidden behind a paywall how about every advertisement’s cpc and cpv goes to user instead of x and user spends that as $0.01 per like and per bookmark and it all goes to the creator? Eg. 100K likes mean $1K to creator no questions asked.