Ben Winston

942 posts

Ben Winston banner
Ben Winston

Ben Winston

@b_winston

Live smart and be kind.

United States Katılım Temmuz 2010
164 Takip Edilen57 Takipçiler
Ben Winston
Ben Winston@b_winston·
It really is remarkable how much more expensive Claude Opus is than using Codex or Gemini. I can easily spend $100-300/day in Opus usage while only spending $2-10/day in Gemini or $4-20/day on Codex. Codex doing higher quality coding at 15x+ cheaper cost is really impressive.
English
0
0
0
26
Ben Winston retweetledi
Emmett Shear
Emmett Shear@eshear·
The Civilization turn loop where you check in on each city as it finishes production, reallocate citizens, and choose the next task is…eerily similar to managing coding agents. Right down to how eventually you just turn on governors and accept the inefficiency.
English
24
31
646
39.7K
Ben Winston
Ben Winston@b_winston·
@paulg Appreciate you keeping the attention where it needs to be 🫡
English
0
0
16
2K
Ben Winston retweetledi
Eren Bali
Eren Bali@erenbali·
Every technical founder who had stopped coding 10 years ago
Eren Bali tweet media
English
183
355
7.9K
533K
Ben Winston
Ben Winston@b_winston·
There's people who want to argue and there's people who want to build, and I vastly prefer the latter. It's so much more enjoyable to keep moving forward rather than constantly try to stop things.
English
0
0
1
16
Ben Winston
Ben Winston@b_winston·
@Noahpinion Losing all the tools I take for granted every day now would be incredibly rough. 80% of my favorite physical objects were a decade from being created too.
English
1
0
5
1.3K
Noah Smith 🐇🇺🇸🇺🇦🇹🇼
If you could reset the world to 2006, and live in that world instead of this one (keeping your current age unchanged), would you do it?
English
108
3
76
59K
Ben Winston
Ben Winston@b_winston·
@mckaywrigley @claudeai yeah, I'm immediately putting this into testing with my OpenClaw agent. It'll be great across a ton of other tasks if true, but with how many tokens my agent burns a day here's hoping this is a truly adequate replacement.
English
0
0
2
1.4K
Mckay Wrigley
Mckay Wrigley@mckaywrigley·
@claudeai i am extremely here for a version of sonnet that is better than opus 4.5
English
9
1
278
32.3K
Claude
Claude@claudeai·
This is Claude Sonnet 4.6: our most capable Sonnet model yet. It’s a full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design. It also features a 1M token context window in beta.
English
1.1K
2.5K
22.2K
7.6M
Ben Winston
Ben Winston@b_winston·
Looking forward to trying this out as a possible replacement for Opus 4.5 in OpenClaw. If we can get significantly cheaper while retaining high performance, it'll be a big boost to potential adoption. For me it means I'll just get to do 5x as much :)
Claude@claudeai

This is Claude Sonnet 4.6: our most capable Sonnet model yet. It’s a full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design. It also features a 1M token context window in beta.

English
0
0
0
21
Ben Winston
Ben Winston@b_winston·
things just keep moving faster and faster!
English
0
0
0
6
Ben Winston
Ben Winston@b_winston·
now this is getting really fun :)
Ben Winston tweet media
English
0
0
0
2
Ben Winston
Ben Winston@b_winston·
There's an interesting phenomenon happening right now where the *best* coders don't write code any more, while lesser coders staunchly avoid using AI heavily. I think most people expected the opposite to happen, and it's kinda wild to see how much different the reality is.
English
0
0
0
5
Ben Winston
Ben Winston@b_winston·
I feel this so much
Ivan Boroja@ivanboroja

Only 3 times in my life i had this "I need to go to to sleep so i can wake up earlier and get back to it" feeling - First time was at about 10 years old, when i got my own pc, i was exploring every nook and cranny - Second time was when i was 23 and was planning on starting a company - Third time is now.. I fell for the hype, bought a mac mini, installed openclaw and boom... two days gone, disappeared just like that I'm having way too much fun, and in a single session i managed to automate a project manager role that would've otherwise cost $75k a year. I have an AI agent running on a cron job, it checks slack, discord and linear every 4-5 hours, scans the channels and flags new tasks, then runs a compatibility algorithm ( fancy word for saying it figures out which team member is the appropriate for that task ) then adds that task to a notion board, and uses a discord webhook to ping team discord with the task. It also breaks down the client's request and suggests a plan of action. I fed it our 90 page internal document on operations so it knows exactly what sort of requests to flag and how much time they would take. I made a bot so now people can interact with it, ask any comapny related questions, added a quirky personality, and already got " i wish it was a real person " comment form a team member ( good thing i spent 6 hours making that personality document ) Mind you, i have basically no coding knowledge, i studied IT in university but dropped out because code was so boring to look at ( i might get some hate comments for that one ), but what i mean to say is that someone like me was able to overnight basically create something like this, and it runs, it fucking works. It's nuts. Excuse my French, i feel like something is coming and will have a brilliant effect on humanity, or doom us all, skynet and all. Either way we're already shifting internally, expanding operations to move in the AI services direction, as Artasaka will be the service company of the future, and now more than ever i believe we're at the right time in history to create something incredible.

English
0
0
1
17
Ben Winston
Ben Winston@b_winston·
@randallb @Austen I'm thinking of this paper where they found "executing...workflows with a single agent matches or slightly exceeds their multi-agent counterparts" and "Single-agent execution is substantially more efficient and cheaper than multi-agent execution." arxiv.org/html/2601.1230…
English
1
0
1
117
Randall Bennett
Randall Bennett@randallb·
@b_winston @Austen This is not a great read on context windows imo. More context for a task = great. More unrelated context = terrible. That's the case for agents / subagents / everything. The right context discovered at precisely the right time is the key, not the length of the context window.
English
2
0
0
184
Austen Allred
Austen Allred@Austen·
A couple downsides of OpenClaw we may be the only people who care about: 1. Subagents can’t spawn subagents by themselves 2. Cron jobs can’t spin be run by subagents directly, only by main orchestrator agent Might fix for ourselves but not sure if anyone else cares
English
49
3
123
9.4K
Ben Winston retweetledi
Andrej Karpathy
Andrej Karpathy@karpathy·
New art project. Train and inference GPT in 243 lines of pure, dependency-free Python. This is the *full* algorithmic content of what is needed. Everything else is just for efficiency. I cannot simplify this any further. gist.github.com/karpathy/8627f…
English
652
3.1K
25.1K
5.2M
Ben Winston
Ben Winston@b_winston·
@Austen The bigger question to me is when this will be pointless because context windows are finally big enough to hold everything in the main. Last I saw the research says sub-agents perform worse than everything running through a single agent if it has large enough context window.
English
5
0
1
8.5K
Ben Winston
Ben Winston@b_winston·
@Austen 1. Yes. 2. I have cron jobs that run like sub-agents with separate models and tasks. Not sure if that's what you're meaning, but I talked with my agent about it below:
Ben Winston tweet media
English
1
0
1
336
Ben Winston
Ben Winston@b_winston·
@Austen software is just more fun when you can create it on the fly :)
English
1
0
0
458
Ben Winston
Ben Winston@b_winston·
AI is still very bad at design. I understand all the reasons why work is being done on coding first, but I'd love to see an actually great AI designer this year.
English
0
0
0
14