Joady

I think the assumption you're making is that the TPUs are going to be or are close to being equivalent to the GPUs from Nvidia and that Nvidia won't accelerate and isn't accelerating even faster. I'm not sure that's entirely true. I think for inference they will definitely use their chips. I think for training it's Nvidia for a very long time and Nvidia is accelerating so hard right now. It's going to be difficult for new technology that is not as generalised as Nvidia's GPUs to actually keep up. A big element is going to be capability and energy efficiency and all the intangibles like interconnects, racking, cooling, etc., and how they can manage those. All of that is possible but I would think it's going to be quite a while before anybody gets there and only Google is even in the ballpark of getting there in the medium term.

Google spent a decade building TPUs before anyone took them seriously. Amazon built Trainium while still spending billions on NVIDIA. Meta built MTIA while running the world's largest social media inference workload on someone else's chips. None of them announced they were leaving NVIDIA. They just quietly built the exit while they still needed the door open. Three of the biggest GPU customers on earth now run production AI workloads on hardware they designed themselves. That's not a threat to NVIDIA's future. It's already happening to their present. The question was never if big tech could beat NVIDIA. It was always how long NVIDIA could keep selling to customers who were building their replacement.

@adampatricknc @RoundtableSpace Maybe that explains why Google has already delayed the release

@RoundtableSpace Will believe this when I hear the US government is delaying the release...

Gemini 3.5 Pro benchmark leak just dropped and the numbers are turning heads. > Reportedly outperforming Claude Fable 5 and GPT-5.6 in internal evals > Significant zero-shot performance improvements over 3.1 Pro > Currently in private validation and testing - Public rollout preparations appear to be underway - July 17 is the targeted launch date If the leaked numbers hold, this isn't Google catching up ,it's taking the lead. Fable 5 just became the model to beat. Gemini 3.5 Pro might already have beaten it.

The mice used in these experiments are genetically altered to be as similar as possible to humans. They're not humans but this is a very promising sign. That doesn't mean you won't have to tweak the formula to make it work in humans but the probability is that it will have some efficacy. I don't think sneering at potential breakthroughs is how you make friends and influence people

@kimmonismus I'd hate to be the one to say it, but you obviously don't know yet: a mouse is not a human. There you go. That's news to you, I know. Take your time to process that fact of life.

One dose of a frog-gut bacterium completely eliminated colorectal tumors in every treated mouse. Not merely shrank them. Complete response. The bacterium, Ewingella americana, multiplied roughly 3,000-fold inside the tumors within 24 hours. It attacked cancer cells directly while recruiting T cells, B cells, and neutrophils. In the experiment, it outperformed four doses of anti-PD-L1 immunotherapy and liposomal doxorubicin. Then researchers rechallenged the cured mice with the tumor: 0/10 developed tumors. 10/10 untreated mice did. The bacterium disappeared from the bloodstream within 24 hours and wasn’t detected in healthy organs. It’s one small mouse study, not a human cancer cure. But the concept is remarkable: a living drug that finds the tumor, multiplies inside it, destroys it, and potentially teaches the immune system to remember. In the foreseeable future, we will cure all cancers.

@bindureddy Following your other tweets: You ranked it where Aritificial Analysis did, ahead of GLM-5.2 but behind Opus 4.8. Did you change your mind?

@Joadys yeah, benchmarks are not real world performance. Please follow my other tweets

Grok 4.5 is a pretty good replacement for Haiku and is actually usable in the real world Very glad to see Grok being useful 🚀

I and almost everyone else I've seen using this feels like it's a faster Opus class model. Your mileage may vary but if you're seeing a Haiku class model (~50% lower all round than Opus), this may be how you're evaluating it rather than the model. A lot of non public benchmarks are also ranking it very highly which is a strong signal of model strength.

@BoyGeorge It's not an airport, you don't have to announce your departure

This is my last post on this site! Good riddance!

@kimmonismus Option 4 They introduce the $500 / month tier with Fable

Three options: - They release Opus 5 on July 19th or 21st to appease the public and remove Fable 5. - They extend Fable 5's support again, and many people make fun of the fact that Anthropic can't withstand OpenAI's pressure. - They find a final solution by the end of next week to ensure Fable 5's long-term viability (more computing power). I now consider it highly unlikely that virtually nothing will happen and Fable 5 will be removed without a solution.

This is the way

i'm glad elon bought cursor so they can actually compete i'm glad openai still has teeth and infinite compute i'm glad anthropic is pushing the frontier and competition forces them not to hide it i'm glad the zuck is taking this seriously i hope google gets its shit together i'm glad china is catching up i'm sad europe is not i don't want concentration of power i want a multipolar world with healthy competition i want humanity to thrive i want my children to grow in a better world i want to cure cancer i want to understand the universe i want people to be truly free, and freedom is just the capacity of execute your own will i want the future to be bright

Hot take, you generally favour Chinese models over American. That's fine. But Grok has come in significantly above any Chinese model so far. Don't let Fandom blind you to the truth. Grok is Opus 4.8 level or better on most serious benchmarks and higher in many private (not overfitted) benchmarks

Hot take about Grok-4.5: positively: It’s a budget-friendly model that's a bit cheaper but slightly less powerful than US frontier AI. negatively: It’s an overpriced model that performs about the same as Chinese AI but costs more.

@MatthewBerman GPT-4.5 A great model at some things like creative writing, but too big a model according to rumour so no-one could afford it

What are other examples in tech of building the best version of something but it costing too much to serve/sell?

I think you're 100% right that coders genuinely will switch to Grok 4.5/Grok Build especially if Fable gets removed from the subscription. Grok 4.5 is much more Claude like that Codex is and that will appeal to many Claude users. I think normies will encounter Grok via X as it's build right in. Whether they switch in numbers remains to be seen.

@Joadys see below and i dont think any normie switches to grok at all coders will though x.com/robinebers/sta…

it’s not “claude vs codex” the real fight is “normie vs engineer” while not universal: engineers tend to prefer codex our brains are wired differently almost everyone else leans towards claude models because that’s how their brains work just ask normies how much they like talking to chatgpt vs claude (news flash: they prefer claude big time) and that’s a problem, because engineers are quickly losing market share yes we’re still ahead, and will be for a while, but just think about how much ground we’ve already lost the real fight is now fought over the everyday people. that’s why you even see Cursor now go after the same non-engineering crowd there are many exceptions of course but as a general rule i’ve believed for a while that this is true

Anthropics big ticket customers are B2B primarily. But Snubbing the subscription users is still a problem because the vast majority of the online folks who influence the AI narative are solo or small company devs who use the subscriptions and are not paying direct API prices like corporations are forced to pay. That's especially important because the vast majority of users or potential users including those in big corporations don't actually know what the best model is. It's actually incredibly hard to know given that there's so much benchmaxxing and overfitting and many of the qualities that make Claude special, show up less in benchmarks and more in real world use. However, if all the social media AI opinion makers who primarily use subscriptions cancel because Fable is omitted, they will become evangelists for another path like Codex, Cursor or Grok Build (It seems like 4.5 is the most claude like experience of the remaining options) Suddenly corportate CTO's who decidde what to spend their AI compute budget on will be seeing other names lionised every day as the "Greatest AI model/dev harness off all time". It won't be long before the trickle down shift away from Claude gathers steam. The subscription users are the canary in the coalmine, the pied piper of hamlin, whose lead and zeitgeist the corporations follow. Snubbing them on subscriptions would remove Anthropics most valuable marketing tool in one fell swoop. I suspect they will release Claude 5 instead but I don't think that will be enough to stop the erosion of reputation. If Fable is indeed not viable economically to serve on the $200 plan, I would far prefer a $500 plan that included it. We would grumble, but most would still pay (especially if limits were higher). It's a delicate dance for them to walk and I'm not at all sure they will manage it.

@VraserX It's already getting impossible to build data centres because they're being protested everywhere. Even more than the cost and the free electricity is just the right to build them. These data centres are two to three years away max

I’m team Sam on this one. Elon calling people scammers while pitching space datacenters is just way too funny. Sam cooked him. 🤣

@Ziva_vibes3y 44111111 Remove the cross pieces from the third and fourth 4s, and add them as two more 1s at the end

Move two sticks to make the largest number possible What's your final total..?🧠

@0xrwu I think it's awesome for coding. I don't know why you're discounting it.

I think Grok 4.5 will 0-1 fast track to being the most dominant model (by raw token usage) for basically anything but coding. At this point, anything Grok 4.5 or above is smart enough to do tool calling for knowledge work (eg go look at our CRM, identify issues, and queue up follow-ups for the BD reps). The limiting factor here is now just getting all the context. It's cheaper(!) than GLM 5.2 AND smarter.

@Yuchenj_UW Of course it makes sense. You have to use the absolute best because even a small drop in productivity or a final output quality is a huge cost to a business

As coding performance converges across AI labs and model sizes, does it still make sense to pay 5x to 10x more for a tiny bit of extra intelligence? Take GLM-5.2 vs. Fable 5. In 95% of tasks, most users cannot tell the difference. But they will definitely notice the bill.

@thdxr How do you get to using 18 times more Sol than Fable and yet Fable makes up 30%? Fable is less than twice as expensive as Sol is via API. Even with 100% output tokens, that wouldn't math out. Are the OpenAI tokens coming from subscriptions?

our team's first week of having both sol and fable the crazy thing is 30% of our cost was on fable