Intelligence is the new Capital. Own it. š Sovereign AI, open source AI, local AI, personified agents, hybrid human + AI orgs ā¢ Serial startup founder
SF Bay Area
Joined October 2008
- Tweets1.4K
- Following427
- Followers831
- Likes2.6K
This is why as much as I love playing around with Grok Bot and Muse, I can't bring myself to run anything critical on them. You just can't have critical jobs dead in the water for 3 days until the next weekly reset due to a sudden limit change.
This goes for personal stuff but even more for startups.
For the important jobs, I'm still using Hermes and OpenClaw. Qwen 3.8 Flash Next and DeepSeek V4.1 Flash 1017 on local GPU / DGX Spark can handle most tasks with 100% predictable capacity, and I can point to whichever frontier subscription I want (or has usage left) as needed.
Of course I do also still use Codex, Claude Code, Grok Build for projects, specialist agents and adversarial review.
While you don't need a fully sovereign AI stack, staying model and harness-agnostic ensures you are always in control.
Well, this kind of blew up. It seems I am not alone in my displeasure with how @elonmusk and @xai handled the Grok @bot usage nerf.
Tin foil hat theory: They'll make more money renting compute to @AnthropicAI than on paying Grok Bot users - so they're fine with the fallout.
Came to the same conclusion: 512gb M5 Ultra Mac Studio is still going to be too slow to want to use for the models it can run, esp. vs. private, US-based APIs.
And it doesn't get you into the next level of models beyond 2x DGX Spark - ie. it still can't run Kimi 3.1 Pro.
This is so tempting, and the delivery dates are not bad, but I have lately been felling much less bullish on this particular model.
512GB on the surface sounds great, but presumably you'd be wanting to run some of the largest available models on it.
This is problematic for me because as good as the memory bandwidth now is on these, running higher active parameter models on this device is probably going to be painfully slow.
I think the mac is much more suited to run models like Qwen 3.8 Next (which has 6B active params) or maybe GLM 5.3 Flash (18b params), which you dont need 512GB for.
The extra memory is useful, but its usefulness does not scale linearly with price, and so I'm just not sure.
Perhaps with some mixed compute solution (DGX Sparks for prefill/computation and M5U for decode) you could have the best of both worlds.
I just dont think I can live with <60toks/s, and certainly not 20.
Glad I passed on the iPhone Duo. Mostly because of the camera quality - but by the time apps are optimized, Duo 2 will already be around the corner. More looking forward to the iPhone XX (20th Anniversary) though. For now, you're better off with an iPad mini.
The reality is food and grocery ordering is already so optimized by the mobile apps that I can't see agents being of value here. It's hard to beat a few taps over 30 seconds to do a re-order. Once tokens stop getting subsidized, no one is going to be willing to pay for an agent to do it slower. Plus no anxiety about the bot doing the wrong thing, stalling, getting overcharged, etc.
I'm bullish on consumer agents generally, but this isn't the killer use case.
I think Hark is a fantastic product, but this free pizza thing was a fail for meā¦ it took 20 mins to orderā¦ if you look at the comments on the post there are a lot of complaints. FWIW this was the first time hark hasnāt been a great experience.
Reminds me a bit of when on-demand economy startups offered freebies that overwhelmed the system and gave a bad first impression.
Spoonrocket offered $1 meals, Flywheel offered $10 rides with no surge pricing on new years. I tried both but had such a bad experience I never went back.
Itās not obvious to me why Hark failed here! But if you do one of these you really need to be prepared for the influx of traffic. And if you try hark here and have a bad experience donāt delete the app- itās actually very good!
Persistent agent teammates should be first-class organizational actors and have the same affordances (and safeguards) as human actors.
Today weāre taking another step toward our vision of an agent for every person and every process.
Last month, we started rolling out Autopilot. Today, Copilot Studio is taking the next step with persistent agents in Microsoft 365. This means each agent can have its own account, so people can email it, add it to a Teams chat or channel or mention it in Word, Excel or PowerPoint. Business process memory carries all the teamwork and context forward in the next cycles all while IT stays in control.
The first capabilities begin rolling out in preview today, with more planned.
aka.ms/persistentagents/blog
One more year like the last and we'll have Fable / Astra level intelligence running 24/7 on a single local GPU or MacBook Pro for the cost of electricity. š¤Æ
Imagine the use cases & possibilities that could unlock.
This will reset everyone's expectations on what consumer experience should be like. Americans will soon want dealing with every state, city, school, bank, health insurance co, etc. to be as easy as this.
Hopefully National Design Studio will publish playbooks to help everyone understand how to make this transition.
Chief Design Officer @jgebbia officially unveils America.gov, the new online home for the United States of America
The dots Demo assumes you will completely intertwine your work and personal lives. Will be interesting to see how that plays out long-term:
1) Will you BYO ChatGPT to work (like mobile) or will you use company-provided (like laptop)?
2) Will you have two different sign-ins where contexts are temporarily intermingled during employment? Will everyone trust that?
3) What happens when you leave the company?
4) If BYO, can you easily auto-expense back work tokens to company? Does company have cost controls for BYO users?
I suspect this a truly seamless experience will be tough for many. Eg. companies will want employees to use Copilot at work.
ChatGPT will truly need to become as dominant in the eyes of consumers as iPhone for companies to cave to BYO.
The reality is the AI product that would save me BY FAR the most time would be if Waymo allowed tweens / teens to ride solo.
(Note this is timed saved, not productivity leverage. In reality, with AI you work the same or more amount of time, but you can do infinitely more.)
Everyone is convinced that consumers are gonna be very excited about agents.
A lot surely will; but the tasks that really consume a lot of peopleās time remains manual (picking up the kid, making a meal, fixing the thing,etc)ā¦ the chief-of-stuff concept is invaluable for certain white collar professionals. Not everyone.
WOW! Meta VR glasses may have finally cracked it from a form-factor and price perspective.
DisplayPort over USB-C is huge. XReal Air is still what I use the most.
Vision Pro was the first Apple product I regret buying. And I had fully expected that Quest 3 was going to be the last product Meta VR ever shipped...
M5 Ultra performance more impressive than I expected. For now there are few compelling arguments to go 2x DGX Spark or Strix Halo for local LLMs.
I reviewed the M5 Ultra Mac Studio with 256 GB of RAM.
Itās the dream Mac for local AI agents. And Iāve gone local-only with Hermes.
My story, with LOTS of interactive charts:
macstories.net/stories/m5-ulā¦
As we mourn the tragedy and heroism of those lost 25 years ago on 9/11, this piece stood out today for me. Worth watching.
We will never forget.
DeepSeek V4.1 Flash is a breakthrough.
Alas, it looks too large to run locally (4x Spark / RTX PRO 6000 or 512gb Mac needed).
But with the API prices, it may not make sense to. My math says API token costs work out to about the same equiv as Sol on a Pro subscription plan.
The iPhone Duo looks more compelling than I expected!
Most excited about the Apple Pencil support. I used to carry only an iPad Pro for work, but switched to a MacBook in the age of agents. Kept an iPad mini for Apple Pencil plus a folding bluetooth keyboard for occasional use.
It's possible iPhone Duo replaces my iPhone, iPad mini, and iPad Pro... Pairing it with a folding bluetooth keyboard looks ideal when you really don't need your MacBook.
After buying the newest iPhone every year since the iPhone 4S, I'm still using the 15 Pro Max.
Glad to see Apple making iPhone interesting again!
Wow. Muse built with multi-layered prompt injection defense, sentinel agent, encrypted credentials vault and Confidential VM?
Was not expecting this from Meta.
Muse Confidential VM is coming soon. Your whole VM, including all data and conversations with your Muse is encrypted with a key held only by you. Meta can't access that data. Because we think people should have this option.
This is where it stands today. We'll keep working on it.
security.muse.ai
Meta really pulled off a naming coup with Muse. The fact that it's a noun is key. It can feel more personal & yours.
"I had my Muse do it." ā
vs.
"I had my Instinct do it." or "I had Instinct do it." ā
Alas I don't see Instinct becoming a verb either like "Google it." ("Instinct it.")
Needed soon: fiduciary duty / duty of loyalty concept for personal agents.
You will want assurance the agent is acting in your best interest vs. that of the agent provider, marketplace, advertiser etc.
Esp. given all the private info / access you may share with your agent.
Suspect there will be a lot of hand-wringing over AI benchmarks and whether new ones are needed.
Capability for autonomous ownership of outcomes and trustworthy decision-making seems like a new frontier that isn't being measured well.
This benchmark seems interesting...
Meet E-Commerce Bench, a new benchmark for long-horizon autonomous business operations. š
Agents start with Ā„100,000 to run online stores for 365 days, handling sourcing, negotiation, pricing, promotions, inventory and cash flow, in a market driven by real e-commerce data.
What's inside: š
- Real economics: 6886 products, 576 suppliers (152 fraudsters), 600-minute workdays, storage fees, returns and reputation.
- Long-horizon learning: almost no model learns to buy cheaper or improve its strategies over a full year of operation.
- Seven-axis evaluation: beyond year-end assets, we score six more dimensions, revealing that no single model dominates across the board.
Learn more about E-Commerce Bench: š
- Blog: qwen.ai/blog?id=e-commerce-bā¦
- Paper: arxiv.org/abs/2608.30730
- Project:Ā ecbench.github.io/
- Code:Ā github.com/QwenLM/E-Commerceā¦
Excited for the future of American open source AI with @NVIDIAAI joining forces with @poolsideai and @huggingface in the space of less than 2 weeks.
Nemotron 3 Nano Omni and 3.5 Lightning already daily drivers. Laguna S 2.1 was for a while, but now all we need is a US open-source LLM in the GLM-5.3 Flash / Qwen 3.8 Flash Next / DeepSeek V4 Flash tier.
Can't wait to see what Nemotron 4 looks like.
Exciting day for NVIDIA and @huggingface.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. They allow every developer, startup, university, industry and country to build with, customize and benefit from AI.
Thank you @ClementDelangue for coming to me.
NVIDIA is going to be a great home for Hugging Face, its community and the future of open models. š¤
blogs.nvidia.com/blog/nvidiaā¦
What a crazy day in open source, local AI. GLM-5.3 (0xAlpha) and Qwen 3.8 Flash Next both land same day. Both runnable on 2x DGX Sparks (and soon 1x Spark or 1 RTX PRO 6000 no doubt).
Sigh... and I'd just finished getting Qwen 3.8 27B to run 150+ tok/sec C1 yesterday on sglang accelerating my local agents 3x!
GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index. At $0.09 Cost per Task, it sits comfortably on the Intelligence vs. Cost per Task Pareto frontier
@Zai_org has released GLM-5.3-Flash, a smaller and cheaper sibling to GLM-5.3 at 320B total parameters and just 18B active parameters. GLM-5.3-Flash supports low/high/max reasoning efforts, and scores 57 evaluation on the Artificial Analysis Intelligence Index with max reasoning effort. This places the model only 3 points behind GLM-5.3 at 60 and in line with GPT-5.6 Terra and Muse Spark 1.2.
On Z AI's first-party API, GLM-5.3-Flash is priced at $0.15 / 1M input tokens and $0.50 / 1M output tokens, just over 10% of the price of GLM-5.3. Cached input tokens are priced at $0.026 / 1M tokens, an 80% discount. Its Cost per Task on the Intelligence Index is $0.09, compared to $0.68 for GLM-5.3 (max), and it sits on the Pareto frontier for Intelligence vs. Cost per Task.
Key results:
ā¤ GLM-5.3-Flash is 3 points behind GLM-5.3 (max) on the Artificial Analysis Intelligence Index, at ~7.5x lower Cost per Task. At $0.09 per Intelligence Index task against $0.68 for GLM-5.3, it sits on the Pareto frontier for Intelligence vs. Cost per Task. It ties GPT-5.6 Terra ($0.51) and Muse Spark 1.2 ($0.40) at 57 while costing ~5.7x and ~4.4x less per task.
ā¤ GLM-5.3-Flash is less token efficient, but its low per-token pricing means this does not translate into a high Cost per Task. The model used 149M output tokens to run the Intelligence Index, ~11% fewer than GLM-5.3 at 168M, but more than Kimi K3 (133M) and Qwen3.8 2.4T A95B (136M) which score the same on the Intelligence Index. Reasoning tokens account for 134M of the 149M total (~90%).
ā¤ GLM-5.3-Flash matches GLM-5.3 on real-world agentic work on GDPval-AA v2. With an Elo of 1770, the model is tied within the margin of error for GLM-5.3 and Grok 4.6. This places it behind only Claude Opus 5 (xhigh and max). On Terminal-Bench v2.1 it also matches GLM-5.3 (84.3% vs 83.9%), and on ĻĀ³-Banking it trails by 3.1 p.p. at 47.2%.
ā¤ GLM-5.3-Flash demonstrates good real-world knowledge and hallucination rate, scoring +7 on AA-Omniscience. Its AA-Omniscience Accuracy is 28%, 6 p.p. below GLM-5.3 (max) at 34% and well below GPT-5.6 Terra at 47%. However, with a Hallucination Rate of 28%, it is an improvement over GLM-5.3 at 30%. In real-world knowledge, GLM-5.3-Flash knows less than the bigger models and frontier proprietary models in its Intelligence Index tier with an accuracy of 28%.
Additional model details:
ā¤ Pricing: On Z AI's first-party API, $0.15 / 1M input tokens and $0.50 / 1M output tokens . Cached input tokens are priced at $0.03/ 1M tokens, an 80% discount.
ā¤ Accessibility: Accessible through Z AI's first-party API at launch.
ā¤ Size: 320B total parameters with 18B active parameters
ā¤ License: MIT
ā¤ Context Window: 400k
