Ody Mbegbu

This joke took me multiple seconds to get. 🤣🤣🤣

@thdxr Not trying to victim blame but Sam in particular is notorious for all the AI existential rhetoric in order to justify massive valuations. Instead of being honest about it. To him and the other CEOs it’s a perception game. But to many people, It’s a real existential crisis

this is really bad i hope this gets everyone to stop and reflect on how poorly this is all going ai is incredible tech that every single person has had at least one incredible experience with for it to be so hated despite that is a huge failure

@zimwha @GergelyOrosz I agree with you. However it means that any model trained on publicly available code is infringing on copyright. I’m not so sure that is a thread Anthropic wants to pull

@GergelyOrosz So the DMCA defense is now just "Your Honor, it's in a different language." International law scholars in ruins.

This is either brilliant or scary: Anthropic accidentally leaked the TS source code of Claude Code (which is closed source). Repos sharing the source are taken down with DMCA. BUT this repo rewrote the code using Python, and so it violates no copyright & cannot be taken down!

@digitalix Honestly? I wouldn’t mind a 64Gb version of that GPU even if it’s above $1000

32GB of VRAM for under $1000! The Intel Arc Pro B70 just landed.

Qwen 3.5 - 27B is very underrated. Its performance blew me away. The 35B constantly underperforms by comparison

@shanselman 8Gb of RAM used to be the standard. And the specs you mentioned is more powerful than a decent PC from 10 years ago. From a computational standpoint, What has really changed from 10 years ago? So you are 100% correct. It’s organizational willpower

The MacBook Neo uses an A18 Chip that is in my iPhone Pro Max, and it runs full macOS competently in eight gigs of RAM I want to plug my iPhone into my thunderbolt dock and run macOS X. It doesn’t seem like it’s a technical problem anymore, now it’s organizational willpower

@t31kx This is precisely why I use opencode. I immediately switch to Gemini 3.1 It goes Claude -> Gemini -> Kimi K2.5 -> Claude

Claude Code: "You've hit your limit · resets 7pm" Me from 5-6.59pm

"We used to debate using tabs vs spaces in code we'd type out"

@rwsamXO @tom_doerr Still doesn’t matter. The swapping has to happen PER TOKEN

@tom_doerr Tried it not worth it at all actually , it takes very long time for just few tokens , but it would be great if there's contributing to make it faster in such way like swapping between ssd and RAM

Runs 405B Llama3.1 on 8GB VRAM github.com/lyogavin/airllm

@HardwareUnboxed I legit think it will succeed but only in very limited narrow scenarios and that realization will tank the global economy

Which one of these do you think will happen? 1. AI Succeeds? 2. AI Fails? Pick carefully… nah don’t worry either option will result in the global economy tanking. My money is on failing, not that it matters.

@amacarthur @josefbender_ @sveltejs @reactjs I know right? Was about to say the same thing. JSX is also weird for crying out loud

@josefbender_ @sveltejs @reactjs Such an odd thing to throw a fit about

You're telling me I should use 3 different characters just to write an if/else statement? No thanks @sveltejs, I am sticking with @reactjs.

You do realize that C compilers are literally part of its training dataset? Did you know that It was trained on the Bible too.. Does that make it a “Man of God” when it quotes scriptures? 🤦🏽‍♂️

@markacianfrani First off the 3090 is pandemic era hardware. There’s more that goes into inference. For example if you spin up multiple agents, the speed difference would flip

I fundamentally do not understand why NVIDIA is supposed to be so much better than Apple Silicon? Comparing 1 3090 to an M3 Max 36GB and it's only like 1 minute faster? Strictly for training?

@vasalex93 @Austen There is also no switching cost for the open source models. Paying $200 to every new frontier model breaks your value proposition

@Austen That hardware will become obsolete in four years and break in 5-8 years. You will never even break even. You also will have a worse experience the entire time. You played yourself.

If you buy $19,000 worth of Mac Studios you can run the local AI models with performance almost as good as what you get from the big providers and skip the $200/mo inference cost

@nickchapsas Gaming is the only thing tying me to Windows

Ok, I'm done with Windows, enough is enough

@nicoloboschi @Kimi_Moonshot The difference isn’t statistically significant. It needs to be more than 5% to actually matter. Real world scenarios aren’t benchmarks and benchmarks have noise in them so yeah they can claim that it’s state of the art

@Kimi_Moonshot maybe I'm stupid but this is not state of the art coding

🥝 Meet Kimi K2.5, Open-Source Visual Agentic Intelligence. 🔹 Global SOTA on Agentic Benchmarks: HLE full set (50.2%), BrowseComp (74.9%) 🔹 Open-source SOTA on Vision and Coding: MMMU Pro (78.5%), VideoMMMU (86.6%), SWE-bench Verified (76.8%) 🔹 Code with Taste: turn chats, images & videos into aesthetic websites with expressive motion. 🔹 Agent Swarm (Beta): self-directed agents working in parallel, at scale. Up to 100 sub-agents, 1,500 tool calls, 4.5× faster compared with single-agent setup. - 🥝 K2.5 is now live on kimi.com in chat mode and agent mode. 🥝 K2.5 Agent Swarm in beta for high-tier users. 🥝 For production-grade coding, you can pair K2.5 with Kimi Code: kimi.com/code - 🔗 API: platform.moonshot.ai 🔗 Tech blog: kimi.com/blogs/kimi-k2-… 🔗 Weights & code: huggingface.co/moonshotai/Kim…

@talb_bluecurve @tomwarren What about when other countries ask for the keys? Why does the FBI get to have the Keys and say EU or UAE don’t get to have it? Do you see how messed up it all is?

@talb_bluecurve @tomwarren 😂😂😂 Because FBI aren’t humans and keys don’t leak.

Microsoft has confirmed it will hand over BitLocker encryption recovery keys to the FBI. You definitely should not be accepting Microsoft’s default of storing BitLocker keys in the cloud when signing into Windows with an MS account theverge.com/news/867244/mi…

@shanselman As long as the monkeys aren’t allowed to read the tests, it is an interesting experiment.

An infinite number of monkeys and an infinite number of typewriters will eventually write Shakespeare. I'm running 20 monkeys on 20 instances of GitHub Copilot CLI with a test harness, and they will eventually pass all these tests. Assuming that they don't comment the tests out to game the system.

@NikTek They were scared around the AMD hype machine at the time. Nvidia learnt their lesson and now we have 5080 ~ 4080 super and only 30% faster than the 3090Ti

I don't think Nvidia will ever produce a GPU that was as special as the GTX 1080 Ti