At least the Chinese models are open source, so you don't need to send money to the Chinese government to use them (unlike Grok 4, where you need to send money to Elon Musk)
This. I would never use Grok, and my sympathy for people who do is essentially zero. I haven’t even tried it. No matter how good it is, I would have to give money and data to a neo‑Nazi, which is an absolute no‑go.
And there are just as well people that would have no sympathy for people who seem to think it's ok to call people nazis for baseless and childish reasons.
The guy spends most of his time signal-boosting deeply racist, antisemitic, white supremacist stuff on X. He's obsessed with stuff like "replacement theory" and constantly insists white people must make as many babies as possible to maintain their cultural superiority and avoid being outnumbered by other races. You don't have to believe me. Go check for yourself.
I do. You’re certainly right. I just don’t like reducing the entire political compass to “marxist/smth.” and “nazis”. That both devalues these words and and leaves out any nuance.
Guys like Trump or Putin are not nazis (yet). They do resemble Mussolini on various and often quite deep levels. So fascist would probably be the more correct term.
As for Musk I’m not sure. Drugs and whatever mental issues he’s suffering from likely are distorting the real picture (which might also be even darker).
The dude performs the Nazi-Salute and openly supports neo-Nazis here in Germany. He lies about "white genocide" and sides with racists. Of course he is a neo-Nazi. Full stop.
Why is he licking Israel boot is he's a neo nazi ?
It's almost as if being a piece of shit doesn't immediately make you a nazi, we should move on from ww2 era lingo, new things need new terms. When everyone is a fascist and a nazi no one is, and it weakens the original terms to the point of them being meaningless
There’s a candidate with a literal Nazi symbol tattooed on his chest and not one democrat condemned him or demanded he drop out. Guess his party affiliation.
So let’s just be clear that nobody is playing this fake outrage game anymore.
He did heil a couple of times. And created mechahitler. And lied about "white genocide" in South africa. And called a white supremacist talking point "the actual truth"
We should be careful of labeling people Nazis, but Elon does seem to be playing on the wrong side of that fence.
You literally just shift the window over by to the next token once you reach the max amount of tokens you want for context window, NOT with what you train on, (only limited with memory now)
This has obvious issues since you're now losing information from the now unseen tokens which becomes significant if your context window is small in comparision of the answer/question you're looking at. That's why companies try to give stupidly large context windows. The problem is they're not training on the large context window, they're training on something smaller (2048 and above). Due to how attention is setup, you can train on a small amount of context and extrapolate it to any number of tokens possible since they train via ROPE which trains the model because on words and their offset to the neighboring words. This allows us to effectively x2,x3,x10,x100 the amount of tokens we generate vs train with with some form consistency BUT still cause a lot of issues consistency wise since the model approaches more of a "this was trained on snippets but not the entire thing" situation where it has a notion of the context but not fundamentally the entire combined context
That’s a very basic way to keep the LLM inferring past the context window size (there’s better, smarter ways) but that’s not at all what the question was which is how they train a 2M token length window. My understanding at a basic level is that you need corpuses that are >2M in length for training data which is where the problem comes in for - there’s only so much long form content and it’s swamped by all the smaller stuff. I think there’s probably tricks now but I suspect it’s still largely an open problem.
AFAIK nobody does that. They train on much much shorter text but with use tricks in the position encoding steps that can be extrapolated by the LLMs. Lile ROPE and YARN etc.
I came here just to complain about that :-) All LLMs I used seem to give more weight to things at the beginning of the context window and omit many details. Eg. I tried this simple thing: pasted a friend's and my CV into Gemini and asked it to recommend topics for a joint conference presentation. Results depended greatly on the order of CVs pasted in.
The model will use the full context if it's been designed well, but you can still increase the size of the window on models where it hasn't. It's just pointless. People who don't know much about LLMs will still think "bigger number is better" though.
Most attention implementations can work across an arbitrarily long context.
The limiting factors are typically:
1. Often there are latency/throughput requirements for model serving which become challenging to fulfill at a certain context length.
2. The model has to be _trained_ to use the desired context length, and training becomes prohibitively expensive at larger contexts.
(2) is even a big enough problem that some popular open source models that claim to support large context lengths in fact are trained on smaller ones and use "context length extension" hacks like YaRN to trick the model into working on longer contexts at inference time.
Who here actually uses Grok? It's sad to see Elon's arc but when he doubled down on some of his political ideas he had it coming with the Tesla sales going down and x.ai not taken seriously.
I've always tried to remain apolitical and unbiased but it's hard to overlook who's behind a technology you wanna buy. Not that sama and others are saints either, it's just Elon's very obvious and vocal about it.
It's a shame, really, because Grok is a good model. But Elon promised to open source the previous model and it took them forever to do that with Grok 3. Sorry, but I wanna buy from someone who keeps their promises ("FSD by next year").
I like grok for noncoding stuff. I find it hasn't been tuned for "Safety" (meaning it isn't tuned much for political correctness). It also seems good at making images and stories up well. I run some choose your own adventures stories with my kids through it. We tell it who each of their characters are and what the theme is for the night and grok gives them each a section of story and 4 choices. They also have the option of choosing something different then suggested. We have it so it cycles around the turns for everyone. Works pretty well, and if the kids wanna go dark (preteen boy) grok doesn't mind the violence.
Kinda reminds me of the video game from enders game.
> meaning it isn't tuned much for political correctness
Is being tuned for right wing viewpoints the same as not being tuned for political correctness? Because there is tuning happening to a specific viewpoint:
Going off OpenRouter's rankings (https://openrouter.ai/rankings), Grok Code Fast 1 is the most used model by a significant margin, and since those metrics are calculated as of this week, that's after providers stopped giving free promotional access to it. Grok 4 Fast is #5 on that list which was never free.
In terms of models, Grok 4 Fast has essentially zero restrictions on safety, which a) makes it unusable for most applications that allow user input and b) makes it extremely useful for certain applications.
In my experience Grok Fast is the best "cheaper" model out there. Far better than Haiku 4.5 and Gemini Flash. I don't think the other cheaper models should be treated seriously at this point.
Gemini Flash is the first model I disable in any tool I use. It's a joke, and to add salt to injury, google announced a "lite" version of that as well!
For at least the last year, I've been using Grok for 90% of my queries. I pay for their $30 plan as well as $20 for Claude Code, which I only use for simple development projects. For anything more complicated, Grok's expert mode has consistently better results.
I throw all my queries at Grok 4 Expert, GPT 5 Thinking and Opus 4.1 Extended Thinking.. for Golang it's been my experience that Grok produce the best results about 90% of the time as well.
I do! I have felt bad vibes from OpenAI for a while now, and eventually defaulted to Grok as somewhat the lesser of many evils. I respect anybody who doesn't wish to use it, but it's good enough for what I need it for. Case in point: it just spit out valid OpenSCAD code for an adapter piece I want to 3D print.
I find it funny that people are still calling Grok "mechahitler" as if that weren't prompted by trolls and the AI model is going to set up concentration camps on every block.
I feel compelled to point out that the Mechahitler thing was prompted by bad actors hiding invisible tokens in tweets, but sure, it's maybe an unpopular opinion.
Basically, the major free options out there for LLMs are OpenAI, Google, Perplexity, DeepSeek, Meta, and Grok. (I could be missing stuff here, but those are the main players.) DeepSeek is out because of China ties. OpenAI and Perplexity have CEOs that seem incredibly shifty to me. I refuse to give Meta and Google any more info than I have to, so I'm avoiding them. Hence we fall back to Grok. Again, maybe not a completely logical progression, but it's my choice and I get to live with the consequences :)
I don't think you can compare the usual internal backstabbing between executives with someone who literally directed and participated in acts of the US Government, and keep saying and doing things to help and nurture a certain side of the political spectrum.
I used Grok to successfully split a large 10K-line file of spaghetti code into multiple smaller well organised files. This was after giving the same task to Claude, OpenAI, and Gemini, all of which consistently failed.
Grok certainly has its uses, but I default to OpenAI for most business tasks and Claude for code.
I've been occasionally using Grok and found it good for devops stuff; specifically it often is able to explain and produce working configurations without getting lost or introducing subtle mistakes as I've sometimes seen with other models.
I have try it a few times in Copilot as code fast 1 because it was advertised. It has never correctly done something so far. Maybe because it's the fast ver ?
Because some tools (AFAIR Kilo Code but I might be wrong) gave it away for free. The model itself was (still is?) free for a while, so I'm not surprised.
For complex refactors, I use "max mode" in Cursor, which in my experience noticeably improves the AI's performance and makes it go for a lot longer before it starts to drift. I haven't looked into how it works exactly, but it works well if you don't mind the extra cost.
I not an expert ai user (and have never touched Codex), but anything remotely important I do, I force the smallest context window possible. I just did something very beautiful using that principle, which will soon be ready to show the world. It would have been a garbled pile of garbage with long context windows.
Obviously major architectural changes need a bigger context window. But try to aggressively modularize your tasks as much as you can, and where possible run batch jobs to keep your workflow moving while each task stays a smaller chunk.
With the current crop of LLMs/agents, I find that refactors still have to be done at a granular level. "I want to make X change. Give me the plan and do not implement it yet. Do the first thing. Do the second thing. Now update the first call site to use the new pattern. You did it wrong and I fixed it in an editor; update the second call site to match the final implementation in $file. Now do the next one. Do the next one. Continue. Continue.", etc.
This post really has no reason to be flagged. I know Elon is controversial, and I have a lot of gripes with his business practices myself, but this is literally just documentation for a frontier LLM. Can we stay on topic?
This. I wouldn't pay to use it, but big context windows are amazing for programming and especially prototyping when you can keep whole codebase in context.
I'd bet 200% the opposite. That the forces & believers of Musk all 400% know that any discussion of Musk is going to look awful for him.
People who flag don't do it because they don't want to dig in. They are almost universally a force for suppression & ignorance, the billionaire imperialist fatcat friend who is desperate to minimize the public eye.
Seems reductive. Some applications require higher context length or fast tokens/s. Consider it a multidimensional Pareto frontier you can optimize for.
Depends. For coding at least, you can divide tasks into high-intelligence ($$$) and low-intelligence ($) tasks. Being able to do low-intelligence tasks super fast and cheap would be quite beneficial. A majority of code edits would fall into the fast-and-cheap subset.
Grok, no matter how good the technology, is just tainted by Elon. It's sad.
Yeah, same for OpenAI because of Sama. I'm proud to say I haven't touched either for over a year. There are enough good alternatives out there.
Not defending anyone, but by that logic, you shouldn’t be using any Chinese models either.
At least the Chinese models are open source, so you don't need to send money to the Chinese government to use them (unlike Grok 4, where you need to send money to Elon Musk)
Why is that? Does Chinese company really equal Chinese government?
lol. is that really a question? even for american companies look who's at the board of those companies...
Yes, and do you believe they are?
This. I would never use Grok, and my sympathy for people who do is essentially zero. I haven’t even tried it. No matter how good it is, I would have to give money and data to a neo‑Nazi, which is an absolute no‑go.
And there are just as well people that would have no sympathy for people who seem to think it's ok to call people nazis for baseless and childish reasons.
Come on, I would also never use Grok because Elon is an arsehole. He's not a neonazi though. That's ridiculous.
Does the internet move so fast that people have forgotten his Roman salute at the trump rally?
That proves he’s an edgy, drugged out jerk. Raising your hand in a certain way is not sufficient to be a nazi..
The guy spends most of his time signal-boosting deeply racist, antisemitic, white supremacist stuff on X. He's obsessed with stuff like "replacement theory" and constantly insists white people must make as many babies as possible to maintain their cultural superiority and avoid being outnumbered by other races. You don't have to believe me. Go check for yourself.
I do. You’re certainly right. I just don’t like reducing the entire political compass to “marxist/smth.” and “nazis”. That both devalues these words and and leaves out any nuance.
Guys like Trump or Putin are not nazis (yet). They do resemble Mussolini on various and often quite deep levels. So fascist would probably be the more correct term.
As for Musk I’m not sure. Drugs and whatever mental issues he’s suffering from likely are distorting the real picture (which might also be even darker).
The dude performs the Nazi-Salute and openly supports neo-Nazis here in Germany. He lies about "white genocide" and sides with racists. Of course he is a neo-Nazi. Full stop.
Why is he licking Israel boot is he's a neo nazi ?
It's almost as if being a piece of shit doesn't immediately make you a nazi, we should move on from ww2 era lingo, new things need new terms. When everyone is a fascist and a nazi no one is, and it weakens the original terms to the point of them being meaningless
There’s a candidate with a literal Nazi symbol tattooed on his chest and not one democrat condemned him or demanded he drop out. Guess his party affiliation.
So let’s just be clear that nobody is playing this fake outrage game anymore.
citation needed
NOOOO! Having a nazi death head tattooed on your body is because he didn’t known what it was! How dare you
Rule of goats
He did heil a couple of times. And created mechahitler. And lied about "white genocide" in South africa. And called a white supremacist talking point "the actual truth"
We should be careful of labeling people Nazis, but Elon does seem to be playing on the wrong side of that fence.
Central European here. If it quaks like a nazi that's exactly what we would call a neo nazi.
Anyone can make a long context window. The key is if your model can make effective use of it or not.
no one makes effective use of long context.
How do they make the context window longer? (serious question, I want to learn how this works)
You literally just shift the window over by to the next token once you reach the max amount of tokens you want for context window, NOT with what you train on, (only limited with memory now)
This has obvious issues since you're now losing information from the now unseen tokens which becomes significant if your context window is small in comparision of the answer/question you're looking at. That's why companies try to give stupidly large context windows. The problem is they're not training on the large context window, they're training on something smaller (2048 and above). Due to how attention is setup, you can train on a small amount of context and extrapolate it to any number of tokens possible since they train via ROPE which trains the model because on words and their offset to the neighboring words. This allows us to effectively x2,x3,x10,x100 the amount of tokens we generate vs train with with some form consistency BUT still cause a lot of issues consistency wise since the model approaches more of a "this was trained on snippets but not the entire thing" situation where it has a notion of the context but not fundamentally the entire combined context
That’s a very basic way to keep the LLM inferring past the context window size (there’s better, smarter ways) but that’s not at all what the question was which is how they train a 2M token length window. My understanding at a basic level is that you need corpuses that are >2M in length for training data which is where the problem comes in for - there’s only so much long form content and it’s swamped by all the smaller stuff. I think there’s probably tricks now but I suspect it’s still largely an open problem.
AFAIK nobody does that. They train on much much shorter text but with use tricks in the position encoding steps that can be extrapolated by the LLMs. Lile ROPE and YARN etc.
I came here just to complain about that :-) All LLMs I used seem to give more weight to things at the beginning of the context window and omit many details. Eg. I tried this simple thing: pasted a friend's and my CV into Gemini and asked it to recommend topics for a joint conference presentation. Results depended greatly on the order of CVs pasted in.
Long context window = huge amounts of vacant VRAM = our servers are fucking empty
But isn't context window dependent on model architecture and not available VRAM that you can just increase or decrease as you like?
The model will use the full context if it's been designed well, but you can still increase the size of the window on models where it hasn't. It's just pointless. People who don't know much about LLMs will still think "bigger number is better" though.
Most attention implementations can work across an arbitrarily long context.
The limiting factors are typically: 1. Often there are latency/throughput requirements for model serving which become challenging to fulfill at a certain context length. 2. The model has to be _trained_ to use the desired context length, and training becomes prohibitively expensive at larger contexts.
(2) is even a big enough problem that some popular open source models that claim to support large context lengths in fact are trained on smaller ones and use "context length extension" hacks like YaRN to trick the model into working on longer contexts at inference time.
No they can't, it's a N^2 algorithm, just fitting it in the context window is a challenge.
And sure maybe not 2mil of it is usable, but they're reliably pushing the frontier here.
Who here actually uses Grok? It's sad to see Elon's arc but when he doubled down on some of his political ideas he had it coming with the Tesla sales going down and x.ai not taken seriously.
I've always tried to remain apolitical and unbiased but it's hard to overlook who's behind a technology you wanna buy. Not that sama and others are saints either, it's just Elon's very obvious and vocal about it.
It's a shame, really, because Grok is a good model. But Elon promised to open source the previous model and it took them forever to do that with Grok 3. Sorry, but I wanna buy from someone who keeps their promises ("FSD by next year").
I like grok for noncoding stuff. I find it hasn't been tuned for "Safety" (meaning it isn't tuned much for political correctness). It also seems good at making images and stories up well. I run some choose your own adventures stories with my kids through it. We tell it who each of their characters are and what the theme is for the night and grok gives them each a section of story and 4 choices. They also have the option of choosing something different then suggested. We have it so it cycles around the turns for everyone. Works pretty well, and if the kids wanna go dark (preteen boy) grok doesn't mind the violence.
Kinda reminds me of the video game from enders game.
> it isn't tuned much for political correctness
It was tuned to be edgy and annoying though (I mean his general style of speech not necessarily the content).
> meaning it isn't tuned much for political correctness
Is being tuned for right wing viewpoints the same as not being tuned for political correctness? Because there is tuning happening to a specific viewpoint:
https://gizmodo.com/elon-says-hes-working-to-fix-grok-after-...
Yeah, but you can argue that the AI has been biased because of biased training data.
Ultimately every AI is biased based on what you train it on and how you instruct it.
I tend to use LLMs from different companies and personally compare them, and read between the lines.
Going off OpenRouter's rankings (https://openrouter.ai/rankings), Grok Code Fast 1 is the most used model by a significant margin, and since those metrics are calculated as of this week, that's after providers stopped giving free promotional access to it. Grok 4 Fast is #5 on that list which was never free.
In terms of models, Grok 4 Fast has essentially zero restrictions on safety, which a) makes it unusable for most applications that allow user input and b) makes it extremely useful for certain applications.
All propietary AIs are probably biased in some way. I mean, that is the power of them and the reason they're propietary, right?
So I tend to use different LLMs from different providers, personally compare them and read between the lines.
In my experience Grok Fast is the best "cheaper" model out there. Far better than Haiku 4.5 and Gemini Flash. I don't think the other cheaper models should be treated seriously at this point.
Gemini Flash is the first model I disable in any tool I use. It's a joke, and to add salt to injury, google announced a "lite" version of that as well!
For at least the last year, I've been using Grok for 90% of my queries. I pay for their $30 plan as well as $20 for Claude Code, which I only use for simple development projects. For anything more complicated, Grok's expert mode has consistently better results.
You're 110% doing something (or many things) wrong.
I throw all my queries at Grok 4 Expert, GPT 5 Thinking and Opus 4.1 Extended Thinking.. for Golang it's been my experience that Grok produce the best results about 90% of the time as well.
Some simple example:
https://claude.ai/share/6d178173-cdf7-4e50-a467-73ee9f479d56.
https://chatgpt.com/share/69102735-46ac-8012-9cf0-0969585c86....
https://grok.com/share/bGVnYWN5LWNvcHk%3D_54b5f2f1-732e-4372....
I don't use Gemini but haven't been impressed whenever I tried it with GitHub Copilot.
I used to think OpenAI was going to be the Yahoo of the AI wave, but might not even be that, maybe it's the AOL.
And from what it looks like to me Google is preparing to be the Google of the AI wave.
Or maybe the Google Wave of the AI Wave...
I do! I have felt bad vibes from OpenAI for a while now, and eventually defaulted to Grok as somewhat the lesser of many evils. I respect anybody who doesn't wish to use it, but it's good enough for what I need it for. Case in point: it just spit out valid OpenSCAD code for an adapter piece I want to 3D print.
Calling mecahitler the least bad option is absolutely wild.
I find it funny that people are still calling Grok "mechahitler" as if that weren't prompted by trolls and the AI model is going to set up concentration camps on every block.
I feel compelled to point out that the Mechahitler thing was prompted by bad actors hiding invisible tokens in tweets, but sure, it's maybe an unpopular opinion.
Basically, the major free options out there for LLMs are OpenAI, Google, Perplexity, DeepSeek, Meta, and Grok. (I could be missing stuff here, but those are the main players.) DeepSeek is out because of China ties. OpenAI and Perplexity have CEOs that seem incredibly shifty to me. I refuse to give Meta and Google any more info than I have to, so I'm avoiding them. Hence we fall back to Grok. Again, maybe not a completely logical progression, but it's my choice and I get to live with the consequences :)
> CEOs that seem incredibly shifty to me
Yet the next level beyond “incredibly” somehow makes it alright again?
The best ones are out for... Reasons? This seems completely bad faith and honestly really Elon musk fanboyish.
Literally none of this options you listed are that objectionable.
Do what the rest of us do and switch frequently. Don't use mekafurhur and you'll be fine.
As you point out, Sam Altman is not exactly an altar boy: https://fastcompany.co.za/business/2025-11-07-sam-altmans-tr...
I don't think you can compare the usual internal backstabbing between executives with someone who literally directed and participated in acts of the US Government, and keep saying and doing things to help and nurture a certain side of the political spectrum.
Both do both.
Not to an even remotely same degree..
Thought this would be about the whistleblower. They didn't even mention it!
I used Grok to successfully split a large 10K-line file of spaghetti code into multiple smaller well organised files. This was after giving the same task to Claude, OpenAI, and Gemini, all of which consistently failed.
Grok certainly has its uses, but I default to OpenAI for most business tasks and Claude for code.
I've been occasionally using Grok and found it good for devops stuff; specifically it often is able to explain and produce working configurations without getting lost or introducing subtle mistakes as I've sometimes seen with other models.
I don't but only because the model is not satisfying, not because I dislike Tesla
I have try it a few times in Copilot as code fast 1 because it was advertised. It has never correctly done something so far. Maybe because it's the fast ver ?
> I've always tried to remain apolitical and unbiased
Clearly
Grok fast is by far the most used model in openrouter with more than a trillion tokens weekly[1].
[1]: https://openrouter.ai/rankings
Because some tools (AFAIR Kilo Code but I might be wrong) gave it away for free. The model itself was (still is?) free for a while, so I'm not surprised.
Openrouter is not counting tokens used by Kilo or Cline. They have own endpoints.
Yet if you go to the actual model’s page:
https://openrouter.ai/x-ai/grok-code-fast-1
Cline and Kilo code are in the top 3. So how does that work?
It’s considerably cheaper than competing models like 2.5 flash, though. So its not that surprising
What models are better than Grok?
Sonnet-4 and onward, GPT-4 and onward
Half of USA voted for Trump. That should answer “who actually uses Grok”.
I personally use the best tool for the job, which Grok sometimes is.
Trump received 77.3 million votes. Harris received 75 million votes. The US population is about 342 million.
I am not sure why these numbers would matter. He won, obviously, because the majority of voters voted for him.
Which are Americans, Americans who either voted for him and didn't do enough against him.
There is really no excuse to democratically vote for a person like this and let all this bullshit happen.
i didn't
Bluntly: you couldn't pay me to use it.
I had a failed refactor with Codex recently and I am wondering if context window size is the cause.
For complex refactors, I use "max mode" in Cursor, which in my experience noticeably improves the AI's performance and makes it go for a lot longer before it starts to drift. I haven't looked into how it works exactly, but it works well if you don't mind the extra cost.
I not an expert ai user (and have never touched Codex), but anything remotely important I do, I force the smallest context window possible. I just did something very beautiful using that principle, which will soon be ready to show the world. It would have been a garbled pile of garbage with long context windows.
Obviously major architectural changes need a bigger context window. But try to aggressively modularize your tasks as much as you can, and where possible run batch jobs to keep your workflow moving while each task stays a smaller chunk.
With the current crop of LLMs/agents, I find that refactors still have to be done at a granular level. "I want to make X change. Give me the plan and do not implement it yet. Do the first thing. Do the second thing. Now update the first call site to use the new pattern. You did it wrong and I fixed it in an editor; update the second call site to match the final implementation in $file. Now do the next one. Do the next one. Continue. Continue.", etc.
This post really has no reason to be flagged. I know Elon is controversial, and I have a lot of gripes with his business practices myself, but this is literally just documentation for a frontier LLM. Can we stay on topic?
Here are the 5 steps of denial:
1. It's not better.
2. It's better but it doesn't matter.
3. It matters but people wouldn't use it.
4. People use it but I wouldn't use it.
5. Ok, I use it but you sir are racist!
We are at around number 2 and 3.
This. I wouldn't pay to use it, but big context windows are amazing for programming and especially prototyping when you can keep whole codebase in context.
Gemini's 1M is amazing.
It's funny how fast this post is flagged, lol. Have other LLMs or blunt ads got the same treatment on HN?
It's probably because lots of people here resent their difference in personal ideology with Elon Musk.
I'd bet 200% the opposite. That the forces & believers of Musk all 400% know that any discussion of Musk is going to look awful for him.
People who flag don't do it because they don't want to dig in. They are almost universally a force for suppression & ignorance, the billionaire imperialist fatcat friend who is desperate to minimize the public eye.
But for some reason if I load a 400kb file into it... it can't even read the file?! Pffft, whatever elon. Go play with your rockets.
Who gives a shit?
What matter is not context or the recod token/s you get.
But the quality for the model. And it seem Grok pushing the wrong metrics again, after launching fast.
Seems reductive. Some applications require higher context length or fast tokens/s. Consider it a multidimensional Pareto frontier you can optimize for.
Depends. For coding at least, you can divide tasks into high-intelligence ($$$) and low-intelligence ($) tasks. Being able to do low-intelligence tasks super fast and cheap would be quite beneficial. A majority of code edits would fall into the fast-and-cheap subset.