Kydo retweeted
I'm excited to this -- we just piloted out our dev ops (CIs) and cut down our internal costs by 50% and improved performance upto 2x as we have much strong Max and Ultra chips on the network.
125 CI hours done.
Apple CPUs rule the world in terms of performance at such an amazing cost.
If you're looking for strong MacOS (or Linux) machines with SOTA CPUs (+ powerful GPU), hit us up -- for RL, MacOS app builds, let us know -- we got thousands more.
Kydo retweeted
On November 5, 2026, our security team detected and contained unauthorized access to the system that stores the console output (job logs) of CI jobs run on Blacksmith runners. The access was revoked the same day, and a later attempt by the same party to regain access was blocked. Our investigation is ongoing. At this time there is no evidence that customer exposure goes beyond the job-log store.
Based on what we know at this time, the incident impacted job-log data belonging to approximately 2.6% of customer accounts. We have contacted each of these customers directly with the specifics for their organization, and we continue to work with them. We do not believe any other Blacksmith customer needs to take action at this time. If our understanding changes, we will update this notice and notify any affected customers directly.
We are continuing to actively investigate and to harden our systems, and we will publish a full report once that work is complete. Please reach out to [email protected] if you have any further questions.
I think CI prices will drop 50% or more because the CI business faces a completely different calculus today.
CI minutes have grown 8x since 2022, while prices have barely changed. With AI, we're shipping more than ever. More PRs mean CI needs are skyrocketing, and so are our CI bills.
To show just how crazy pricing is today, here's the typical price for macOS CI on a standard 12-CPU instance: $0.16/min.
A new Mac Mini with an M5 Pro (15 CPUs, 48GB) costs $2,299. At $0.16/min, you can recover the cost of that machine in 10 days of continuous use.
10 days.
Running at full capacity for a year, it can generate $84,096. That's a 3,600% ROI...
With this increase in demand, I also expect utilization rates to be much higher than before. That would justify more competition and a lower price point.
Especially if prices stay where they are, I think more people will move to self-hosted solutions, because the ROI on those machines is probably far higher than anything else.
Overall good news for Darkbloom Sandbox as we will bring more supply online.
Kydo retweeted
I am looking to get over to SF/USA next year for a week or two.
I'd love to visit @Apple, @NVIDIAAI, @AMD and @darkbloomai, along with some of the amazing community/contributors over there.
You guys cook like there's no tomorrow, and I am inspired!!
Would anyone from the UK be game to join me?
I have this vision of a bunch of us building something together in 24h that could change the lives of millions for the better 😎
We've been doing one release every day for the past few weeks, and it's getting hard for users to keep track of what's changing. Any recommendations?
life is literally just about the story you tell yourself. you can be depressed that you have a rare genetic condition or think you are a "psa 10 base set first edition holo charizard."
For context, I have several extremely rare hereditary bleeding disorders, including PAI-1 deficiency which has <100 diagnosis worldwide dead or alive
The odds of having all of them simultaneously is like 2.5*10^13 or something stupid like that
My blood is genuinely a PSA 10 base set first edition holo Charizard for hematologists
There are samples of it in unis and research labs all around the world, and I’ve been donating blood for research pretty much my entire life
2/x
kinda surprised i didnt see a fat joke somewhere on the tl today making fun of american portion sizes or sth
Kydo retweeted
We need enough time to gather evidence.
Nowadays teams feel powerful because it’s never been easier to build disposable code.
The issue is, anything outside your core business is a distraction.
Great teams understand it’s about where to deploy (energy + tokens) and standing on the shoulders of giants.
Kydo retweeted
It was marketed as:
- 2 years until permanent underclass
- you’ll all lose your jobs
- your skills will be worthless
- we’re going to replace you
- your entire career is now a $20/month subscription
All in an already terrible job market btw. Possibly the most important technology humanity has ever built, with some of the worst marketing imaginable. The labs completely failed communicating the tech, miserably bad and only improved just end of last year. How did anyone expect that to make people excited?
Kydo retweeted
GM GM!
Quick CUDA.fast and MLX.fast update.
We’ve hit a wall, and need a breakthrough!
CUDA: 179.5%
MLX: 81.0%
We’ve got a little over 24 hours to go…let’s move the needle!
@yukonresearch @ashxhart @TheDavidTai
D/P disaggregation on Mac is really important because many older machines are bottlenecked by prefill speed even tho they have massive RAM.
oMLX 0.7.4.dev1 is out!
github.com/jundot/omlx/relea…
With the new Prefill Accelerator, connect an M5 Max to your big Mac and prefill up to 2.9x faster!
Prefill Accelerator (64K prompt, M3 Ultra 512GB + M5 Max 128GB)
- DeepSeek V4.1 Flash: 1,060 -> 3,123 tok/s (+195%)
- Qwen3.8 Flash-Next: 1,309 -> 3,666 tok/s (+180%)
- GLM-5.3 Flash: 540 -> 1,388 tok/s (+157%)
- MiniMax M3.1: 297 -> 569 tok/s (+92%)
- An M5 Max 64GB still adds 30-118%, and decode speed stays the same.
Faster DeepSeek V4.1 with A8 and buffer pool improvements
- Prefill on M5 Ultra is 43-48% faster, 32K goes 2,631 -> 3,779 tok/s (+44%).
- oQ4e builds now run INT8 Activation Prefill on M5.
Faster Qwen3.8 Flash-Next
- FP16 builds get the fused kernels. Decode on M3 Ultra goes 38.9 -> 63.0 tok/s (+62%), and +23% on M2 Max.
- Decode after long prompts is 3-5% faster on M5 Max.
Better for agents
- Parallel tool results from Claude Code reach the model in call order, so Qwen no longer pairs a result with the wrong call.
- Agent sessions on a cluster reuse the prompt cache across tool turns, 743 -> 9,266 cached tokens by the fifth turn.
- Switch oMLX models inside Codex and Copilot with /model.
Also in 0.7.4.dev1
- Ming-3.0 TTS with a 2x faster decode loop.
- Ask AI can use an oMLX server on another Mac.
- Check for model updates in the macOS app.
- German across the web dashboard and the macOS app.
- Batched decode with 6 requests on 8-bit models is 16% faster.
A big thank you to everyone who sent PRs, reported bugs and tested builds for this release, and a warm welcome to the five new contributors!
Kydo retweeted
Getting a local LLM with looped transformers would really help with compute strong, memory limited systems like HALOs and Sparks. I predict that open model looped transformers, if released, would dominate local inference.
Kydo retweeted
Local inference is likely going to become much more expensive next year.
LPDDR5X is the ram used in every local inference box (Sparks, Halos, Macs) right now and ramping Vera Rubin NVL72 production is competing directly with local LLM consumers.
Each NVL72 contains up to 55% more LPDDR5X compared to a previous generation Blackwell GB300 and 10-30x faster for AI inference and training. The insatiable demand by hyper scalers for frontier inference servers will likely cause a 40% shortfall in LPDDR5X based on open source production and demand numbers.
Some of this gap might be made up from recycling memory chips from GB200 servers but no matter how you slice it, the local inference LPDDR5X market will become much tighter next year.
