AI/Tech Nerd. CTO. Builder. Golf Addict. Shut Up and Ship It. 🇺🇸 I'm having fun. Hope you are too.
Murica
Joined May 2011
- Tweets5.5K
- Following1.2K
- Followers502
- Likes14.7K
Opus 5.5 and DeepSeek V4.1 Flash (via OpenCode Go) executing a plan Fable 5.2 wrote and all of it happening in T3Code. 👨🍳💋
34
I am on edge. There hasn't been a catastrophic wave of CVEs yet this week... I feel like the water is building behind the dam right now...
11
holy smokes this is good
4K image generation in 10 seconds! ⚡
Meet MiniMax H3-Image.
16 million pixel resolution. Stunning realism. Precise prompt following & editing. Up to 16 reference images. All in 10 seconds.
And it's open-weight. 🚀
Available now via our API, MiniMax Code, and Hailuo AI — with 50% off 16MP+ generations during launch week!
Can't wait to see what you create. 🔥
Hugging Face: huggingface.co/MiniMaxAI/Min…
API: platform.minimax.io/docs/gui…
MiniMax Code: code.minimax.io/about
12
@opencode why if I pay $10 do I get $100 of DeepSeek v4.1 but if I pay $40 I only get $200. Help me math?
17
This is insane. Even if a human hand jammed it. But that this is AI generated I am absolutely blown away.
i'm sorry i dont care which model made this but it blows my mind that this is made by an llm AND runs in the browser
claude.ai/artifact/BGY1SztrA…
16
Yes. This is correct.
Don't listen to TLC (🎵Don't Go Chasing Waterfalls🎵).
Since the software development lifecycle has gone from months/years to hours/days, a big upfront plan is again useful and crucial to the success of AI driven software development.
Hear me out:
We hated "waterfall" because after Big Upfront Planning, implementation took FOREVER, and by the time we were done, the plan was out of date
But with AI, implementation is rapid (100x faster, or more, than before)
So Big Upfront Planning can now make sense?
1
1
85
Grok 4.8 this week in Bot.
Grok 4.7 was a disappointment. Elon admitted as much. But I am pumped about the new releases they have planned.
Elon posted a detailed update on Grok 4.8 today (Nov 10).
He said they dumped JAX and SGLang and rewrote the entire RL and inference stack from scratch in C/C++ (mostly C, a little C++, some assembly). That took extra time. Pre-training is done, they cleaned up a lot of the SFT data, and 4.8 is now in RL with the evals starting to climb.
Soft rollout to Grok Bot users later this week.
“Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.” - Elon
The new stack comes from the Starlink engineering work. He noted that at SpaceX they write their own headers and calculate their own time (including leap seconds), and that the Starlink software for moving huge amounts of data through space and to the ground turns out to be well suited to training superintelligence.
The code is tight and was written by humans; AI isn’t good enough yet for that class of high-performance systems software.
When someone asked how 4.8 compares to Opus 5.5, he replied: “Not as good. Opus 5.5 is a masterpiece. But Grok 5 should exceed Opus 5.5 and maybe that happens in 2 or 3 months, with continued acceleration thereafter.”
Earlier (Oct 30) he had said 4.8 “will not be a disappointment” and that they’re not optimizing for benchmark. He cares about it being useful.
Back in September he sketched a rough ladder: 4.7 roughly on par with Opus 5.0, 4.8 a noticeable improvement, 4.9 Astra/Fable class, and Grok 5 maybe better than anything. 4.8 has been described as around 2.5T parameters on the new stack.
The short version is that 4.8 is the clean-slate version with rebuilt foundations, landing in soft form for Bot users this week. It’s not positioned as matching the current frontier yet. The bigger jump is framed for Grok 5 in the next couple of months if the new stack performs as hoped.
44
Tailscale is the way. But beware, make sure your Tailnet is locked down tight.
With the usefulness and user base of Tailscale, I have no doubt threat actors are already trying to compromise them.
we launched @Tailscale support for grok @bot!
let us know what feedback you have for how we can make this better 💗
Help me understand the struggle with "AI Software Factories".
It's simply controlled determistic and non-deterministic loops using different SOTA commercial and open source models.
When the plan is good, this just works.
What am I missing?
1
1
92
Kuberay is elite. Lots of bigtime orgs use it for enterprise scale ML and inferencing.
Big news at KubeCon Americas.
KubeRay is joining the @CloudNativeFdn
Keep your KubeRay stack and add the Anyscale Control Plane in minutes. Find us at #KubeCon booth #P317.
na2.hubs.ly/H08TY0t0
10
This is the way.
BREAKING: Reindustrialization goes nuclear. America approves its first advanced nuclear facility at an industrial site, a major Dow chemical manufacturing plant in Texas.
The Nuclear Regulatory Commission just approved a construction permit for four X-energy Xe-100 reactors, designed to supply both electricity and high-temperature steam to Dow’s chemical manufacturing operations. Chemical production is traditionally a very energy-intensive process, requiring large amounts of electricity and heat.
This marks the first U.S. construction permit for high-temperature gas-cooled reactors in nearly 60 years.
ALT Source: X-Energy
1
64
Exactly what I have been needing
Want your own Slack for $0? I open-sourced Ember. It deploys to @Cloudflare’s free tier in about a minute, works with Slack-compatible payloads, and includes an AI assistant and media attachments.
9
I’m building an enterprise hypervisor to compete with VMWare.
Anyone know someone with idle enterprise hardware that I could get test and build time on for like a week at a time?
1
22
Solid update on Hetzner nodes.
I have run into this before and it took us a while to figure out.
If you’ve ever had a Hetzner Cloud server start dropping new connections under load while CPU/RAM still looked fine, this is for you.
Hetzner’s new hcnet raises the Cloud Firewall connection-tracking limit from 80k shared to 120k IPv4 + 120k IPv6.
Hit the old limit and new connections just got dropped.
Free capacity boost for high-concurrency stuff (APIs, proxies, websockets, etc.).
hcnet, our new eBPF-based cloud network stack, is in beta. Each server gets 120,000 concurrent connections for IPv4 and another 120,000 for IPv6. Before, it was 80,000 for both together.
Heads-up, switching either way briefly interrupts the connection.
1
64