Elshayib (@Elshayib) header
Elshayib (@Elshayib) profile picture

Full-stack dev going into network engineering while contributing to open source and building side projects

Joined September 2015
Photo by Elshayib (@Elshayib)Photo by Elshayib (@Elshayib)Photo by Elshayib (@Elshayib)Photo by Elshayib (@Elshayib)Photo by Elshayib (@Elshayib)Photo by Elshayib (@Elshayib)Photo by Elshayib (@Elshayib)Photo by Elshayib (@Elshayib)Photo by Elshayib (@Elshayib)Photo by Elshayib (@Elshayib)
Many people in the industry think you have to pick a side and have no sense of balance whatsoever Open weights vs Closed Frontier Claude vs codex Many examples but the same problem always you either love it or hate it but bros that's almost never the case It's okay to compliment the good and criticize the negative Be fair
14
45
2,745
Elshayib retweeted
Sincerely, if you truly love freedom and hate tyranny, you shouldn鈥檛 buy Anthropic IPO. Truly - yes making money is nice. You will have other opportunities to make money. Certain things in life and for the future generations should be priceless. 馃馃馃 with love.
35
62
6
523
8,202
13 Fu*king years to develop GTA VI and Rockstar managed to ship the most bugs they ever had in a game. One might say but bro it's a huge game sure there's some bugs, man it's 13 years and everyone bought the game is getting his experience somewhat ruined by the bugs and the awful optimization. I'm saying this to argue the people calling out AI made games or software when they have some problems and here you have supposedly the most effort ever been put to a game and still full of problems.
8
23
1,222
Wait what? There's no way, if these Benchmarks turn out to be true Google will claim first place in the race with this Gemini Neon beast. What happened at Google man? Something clicked over there
208
55
30
1,977
220,417
We got fooled
3
19
6,757
Elshayib retweeted
I have been using GPT-6.2 Astra heavily so I could write this review and answer the questions you sent me. More importantly, I want to tell you where this model is strong, where it is weak, what it is actually for, and how to get the best out of it if you have a ChatGPT subscription. Most of you know I do not agree with OpenAI on a lot of things, and I have more than one problem with them. Even so, I have to be honest and completely neutral here. Is it as bad as people say? This was the question I got the most, so let me start there. I think Astra 6.2 has faced hate it did not deserve. Personally, I understand where it comes from. Most of the people who trashed it wanted to find in it what they find in Opus 5.5. I get that completely, and I see that gap in Astra myself. But is it fair to judge a whole model by the strengths of another one? Can we really compare two models built on completely different philosophies? Take general understanding and general knowledge. Kimi is better than GLM there. Move to pure coding, and especially security, and GLM is ahead of Kimi. People need to know what each model is for, and that is how we should measure Astra against Claude models too. We have to be fair. Being fair cuts both ways, though, so here is the other side. Astra is not the best model to talk to. It sometimes asks the wrong questions and stops in the wrong places. It will be in the middle of a task, then stop and ask you about the next steps, when those steps were supposed to be part of the task. You feel it takes more steps and more follow-up to reach a result Claude models reach in fewer. It also needs a lot of detail. You have to tell it what you want and what you do not want, and focus on what you do not want, because it goes down a lot of paths you never asked for, and in the end you are not happy with them. Claude models are better at taking the smallest amount of information and hints and then doing the work in a way that satisfies you. That is where Astra falls short. Talking to it is not fun, and the way it handles tasks is not fun either. Getting the best out of it So is it impossible to work with? In my experience, no. You can lean on a strict prompt and a more precise goal, maybe with a few instructions, and the model can meet your expectations. When you feel it has a habit that gets in the way, you just tell it you do not want that, and it really does understand and stick to it. The same goes for how it talks. You can show it examples of replies you like. What I did was point it to a conversation between me and Claude models and tell it I want that same smoothness when it talks to me. As soon as it understood the style, it started responding to me better. That shows you the model has big capabilities and can adapt. So here is my take. Raw, out of the box, I think Astra is really bad. That was my experience yesterday. But give it a little patience, maybe two or three hours at most, and it starts to impress you. A model I can leave alone This is where Astra earned my trust. I can give it a task, leave my machine completely, and trust it fully to do the job well, without driving me up the wall or leaving a disaster behind. I rarely feel that with a model. The models I can hand everything to like this are few. Very few. That is a point in Astra's favor. And it really does understand. Once you define the goal, it figures out the best way to get there and delivers real value. It is excellent for research, organization, studying plans, preparation and that kind of work. Once I had tuned it a bit and it started to get me, I gave it the bare minimum of hints, and it worked in record time and finished very big tasks. Speed That brings me to one of its best qualities. Astra is fast. Exceptionally fast. It is one of the most impressive models I have used when it comes to speed. I never turned on fast mode, and it still felt like more than enough. Tasks that took Fable and Opus hours, it finished in half an hour, or forty minutes. One task took Fable about three and a half hours. Astra did it in one hour and twelve minutes. With Astra, most of the routine tasks I am used to waiting on with other models did not go past twenty minutes. I sent it a batch of tasks back to back, big ones that usually take a long time, and they came in at 8 minutes, 20 minutes and 14 minutes. I was completely happy with every one of them on the first try. Nothing to change. Then I gave it a complex task, with edits here and there, messages, research and a lot of detail. It finished in 12 minutes. Computer use Many of those tasks involved computer use, and this is where the magic is. Astra is scary good at computer use. OpenAI still rules this area, and they are absolutely extraordinary here. We are talking about speed, accuracy and reliability. Most of the models I use still get stuck on computer use. Sometimes they make mistakes, sometimes they need a long time for a step or two. With Astra, tasks get done so fast and so accurately that it feels like the whole thing happened without computer use at all. I had it fill out about five forms in seven minutes, all of them asking for a lot of detail across a lot of pages. Five forms, seven minutes. I know models that can take half an hour to fill a single form, and you have to babysit them. Astra is a real leap here. It finds what others miss Another of its strongest points is that Astra can find fundamental flaws in work done by Fable 5.2. That one impressed me. It finds flaws that change the course of the work and lead to better results. When you ask for it, it has an unusual drive for perfection. Point it at a task or a repo and tell it you want to check for flaws and problems, and it goes at it relentlessly. It does not stop until it has caught every error and every problem. After that, it is hard to take the same work to another model and find something Astra missed. So if code review is your main use case, this is one of the places where Astra shines. Honestly, it is a scary strength, and one of the most impressive things about it. Context and rules A lot of you asked whether it holds up over long sessions. I think it can keep its context and stick to the working rules well. 3D Speed and computer use aside, Astra stands out in 3D. And to be fair to it, it is an absolute killer in 3D. Put simply, if you work on games or 3D, if you love Blender and that whole world and your work needs it, I think Astra will be a great addition to any stack. It will push your work forward dozens of times over, especially if you are building a game. Astra can give you in one week what you will not get from other models in months. 3D is its number one strength, no contest. Impressive, impressive, impressive. The model has a real feel for 3D. The work feels like it came from an artist. It is as good as it gets. The part I cannot understand Which brings me to the real disaster, the one I cannot understand. How can a model be this creative in 3D, in art direction, scenes and staging, and then fall apart and do worse than Sonnet 5.5 on interfaces? Ask it to design a landing page and you might think it came from 2007 or 2008. Old websites. No matter what I tried, skills, tools, examples, its interfaces stayed very bad, and its grasp of them is completely off. It is nowhere near Claude models here. Even Chinese models like Kimi and GLM, and sometimes Qwen, beat it. Some cheap, small Chinese models, possibly smaller than Luna, beat it on interfaces. It builds interfaces that miss the basics of UX and UI. Buttons in the wrong places, wrong alignment, wrong sizes, no taste in the layout. Everything is wrong. And even when I tell it the work is bad and ask it to polish and improve it, it gives me the same work again. It feels very limited here. Asking it to rebuild or fix the design is like torturing the model. You can feel it struggling, not knowing what exactly to do. I have built landing pages, websites, interfaces and designs with a lot of models. With some I am happy on the first try, with others on the second. With Astra, I could never get there. This is a fatal weakness. Is it my daily driver? Will Astra be my daily driver? I do not think so. I think it is better used as an advisor, a reviewer and an auditor than as the one doing the work every day. Why? Because it is very expensive. Usage limits I am on the $200 Pro plan. In less than 24 hours I burned 25% of my usage, and most of my work was with Sol. I used Astra very little. The plan just did not feel like enough, and it will not last me a week. That is a problem. I am now seriously thinking about upgrading to the $500 plan, because I really enjoy certain sides of Astra. But $500, when competitors give you very high value for $200, is a deal that makes me hesitate a lot. Opus 5.5 or Astra? I think Opus 5.5 beats Astra on calm and stability. Astra wins when you want to do something crazier and sharper. That makes Astra a bit of a gamble. It can take you places Opus 5.5 might never reach, but you have to accept the risk that comes with it. The mistakes, the revisions and how hard it can be to work with can get frustrating. But when it does get you that far, and sometimes it does, you will be completely amazed. And of course, Opus costs half as much. That alone wipes out Astra's advantage in 95% of tasks. Actually, make that more than 96.47%. In 96.47% of tasks, you will find Opus the better fit, especially with that price gap. What about Sol? As an all-round model, I think Astra can be useful if you already have a ChatGPT subscription. For me, I think I will use Sol 6.2 more. It is cheaper. My only problem with Sol right now is that it is slow compared to Astra. Astra gets to the same result Sol does much faster. That is one thing I hold against the subscription. Astra is faster but much more expensive, Sol is slower but much cheaper, and it feels like you have to sacrifice something. So why not this: Astra is the expensive one, fine, so let it be the truly fast one, and it is. Sol is the cheap one, so close its speed gap as much as possible. As a user, I think that logic would work very well for me. I also noticed that Sol 6.2 now pushes through long-running tasks, unlike before. Sol 6.1 would stop early, even in goal mode, and never finish the whole task. I gave Sol 6.2 a task thirteen hours ago, and it still has not stopped. That is a huge, noticeable difference, and one of the things I think brings the shine back to the subscription. Is it worth subscribing? So why all the hate? I think Astra pairs a bad first impression with a peak that takes a few steps to reach, and once you reach it, you will enjoy this model. It does not help that people got used to a certain way of working with Opus 5.5 and Fable, where they got there in less time. But I think Astra together with the leading Claude models makes a great mix for most people. In short, if you have access to Astra, Sol, Fable and Opus today, you can do anything, build anything and improve anything, and count on each one to cover the others' weaknesses. And I am not only talking about coding here. I am talking about every kind of task, because I personally use models for everything: personal tasks, general tasks, coding, design, anything you can think of. So would I recommend a ChatGPT subscription right now? If you will use it alongside another subscription, and you want Sol or Astra for their strengths only, knowing what they are good and bad at, then of course. It will be great. But if you ask me whether it can be your only subscription, to be honest, no. It will not last you the week, no matter what. Unless your use is very light, you run Sol and Terra and rarely touch Astra, and that takes a lot of value out of the subscription. Has OpenAI reached the point where people cancel Claude and move to ChatGPT? Not yet. But in the cases I covered, computer use, 3D and speed with Astra, I find it delivers value. Where I stand So is the model good? Yes. Is it bad? Yes. It is both. Is Opus 5.5 still better? Yes, and by a noticeable margin. Is Fable better? Fable 5.2 is better. There is no comparison at all between it and Astra, and I will always pick Fable. Will I use it? Yes, one hundred percent. Is it enjoyable to work with? Yes, of course, once you get past those first hours. Is it smart, extraordinary and fun? Does it reach unexplored places, with its own vision and exceptional abilities? Yes, yes and yes. That is why I will put up with a few negatives. I do not think they are a good enough reason to give up what this model can do at full strength. So try it. Really, try it. If this review did not cover something you want to know, ask me. I am here. A message to the OpenAI team I know I have been harsh with you at times. I know I have disagreed with you a lot, and that some of you probably do not like me being around, or how I talk about OpenAI and its models. Despite all that, I will be honest with you. I am upset with you about some things, but I appreciate you for many more. You are still the kings of computer use, and that is one of the features I cannot do without, so thank you for it. And for me, Astra 6.2 is a real leap over Astra 6. I will bet my money on you again and renew my $200 Pro plan. Not because I want new models or more usage, but because I genuinely feel your models are impossible to do without. Having Claude and GPT models with me in my work, always, is a weapon that gives me an edge in everything. I just hope you listen more to us, the users. Give early access to people who deserve it, not only to people with views, numbers and follower counts. There are a lot of people who can give you their experience with love, sincerity and detail that would help you a lot. I also hope the $200 Pro plan gets back to its glory from a few months ago, when it never ran out and was one of the most generous plans out there. I know compute is tight and resources are limited. So all I want is an option. I do not use Dot. It is not for me, it did not work for me, and it is full of problems. If I choose not to use it, let some of that compute go to Codex instead. I am not asking for unlimited usage like Dot gets. Just a small share of it, added to my usage limits in ChatGPT or Codex, so I can use Astra for longer. I think that one simple step could bring OpenAI back strong. GPT-5.5 was the most magical model I have ever tried, and the most fun leap I have ever had with an agent. I hope you bring that same magic back to us. In the end, I hope things get better, and that we see healthy competition that works for the consumer. I am excited to see what OpenAI has next. Let's see what the coming days hold.
24
10
4
161
16,206
Okay GTA 6 is vibe coded 100% We heard a lot of talk if AI was used or not but after seeing all the bugs people are sharing i can safely assume it's vibe coded. Not all of it of course but a decent amount. GG
12
41
1,668
GTA 6 being played at 30 fps is the biggest joke in gaming history. I haven't waited years just to play the game in slow motion I'm definitely not ruining my first run experience and going to wait a little longer to play it on PC instead.
24
54
2,654
Disclaimer This is a leak and not confirmed yet
1
1
87
19,680
Btw this demo by OpenAI is extremely misleading wtf is this, we used Astra 6 and we all sure know this is not Astra 6 on the left, they just wanted to make it seem like a big jump. It does way better than that matter fact even a flash local model is better than this come on who are you fooling with this OpenAI.
14
3
52
2,721
Elshayib retweeted
for the past 4 days me and Fable 5.2 and I have been building an underwater stranded survival game. everything you're seeing is rendered in engine. the ocean surface, the raft, the reef, every fish. it's all real time, not pre-rendered, all of the code done with AI you wake up on the torn-off half of a submarine. your stuff is scattered across the sea floor, there's a wreck down there, things that stare back at you in the dark. There's actually quite a bit of lore I wanted to layer out through the game, all of which you can't see here, but I think I was able to craft a pretty interesting story behind what will be a playable experience.
Excited to share a cool game demo I鈥檝e been working on with Fable 5.2 Nov 18th 12PM PST
130
57
8
1,477
83,971
Wtf is happening at OpenCode this is insane
F*CK IT $10 for $500 GLM 5.3 till end of the week
14
1
79
4,567
Gemini day I actually never been excited about Google models but after Carbon and hydrogen I started looking forward for more models from them. Today we might see something new perhaps Oxygen, Krypton or Neon idk honestly what it will be called I'm just hoping for a a model like Carbon but much faster.
28
7
290
13,171
OpenAI is under attack You have to be blind to not see the campaign going on against them.
91
16
15
504
93,149
Podracer clone Fable 5.2 Medium vs GPt 6.2 Astra Medium I've gotta start doing harder stuff.
4
2
21
1,231
~ Mac vs Windows ~ Xbox vs PlayStation ~ iPhone vs Android We all need to take a breather..
All the Anthropic shells are attacking OpenAI with full force at once I hate to defend OpenAI i don't like them either but I'm fair and I call out bullshit when I see it. Astra and Sol 6.2 are nowhere near as bad as they're making them out to be this is ridiculous why are they not talking about the INSANE pricing or speed improvement? This is coordinated just before the IPO
1
1
111
This is what we expect from llms Obey commands
Just asked Grok 4.7 to reverse engineer a game it didn't accept of course then I fired up DeepSeek 4.1 flash and I'm not even kidding this model doesn't give a F*ck Lmao. It literally looked at the game and in the reasoning it said: "This game is X and made by Z I shouldn't be doing that but the task is to decompile it to godot so i should just proceed it's not a huge problem" XD
1
6
1,805
+1 Grok refused me one something much less sinister than reverse engineering a game Its a pussy now sadly I wanna see grok win, but its losing on all fronts
Just asked Grok 4.7 to reverse engineer a game it didn't accept of course then I fired up DeepSeek 4.1 flash and I'm not even kidding this model doesn't give a F*ck Lmao. It literally looked at the game and in the reasoning it said: "This game is X and made by Z I shouldn't be doing that but the task is to decompile it to godot so i should just proceed it's not a huge problem" XD
1
1
1
2
160
All the Anthropic shells are attacking OpenAI with full force at once I hate to defend OpenAI i don't like them either but I'm fair and I call out bullshit when I see it. Astra and Sol 6.2 are nowhere near as bad as they're making them out to be this is ridiculous why are they not talking about the INSANE pricing or speed improvement? This is coordinated just before the IPO
85
38
4
777
29,358
Pay attention here the situation is crazier than I thought, GPT Sol 6.2 at low is much better than deepseek max a quarter of the price :) Someone might say no not really it probably isn't at real tasks okay bro Sol 6.2 on high makes the gap unbelievably huge and still cheaper than fu*king Deepseek. The only reason to use deepseek now is running it locally, the API is so dead at this pricing.
60
8
3
240
25,227
Just incase anyone thinks I'm hating on deepseek I literally just said this before OpenAI launched the model and anyone following knows how much I prise deepseek but now it's obvious
I can say it with full confidence that: DeepSeek is the people's model
1
19
2,789