Oh how exciting! @RosemaryOrchard joins @inhatko @mikahsargent and me for MacBreak Weekly! We're live now at youtube.com/twit
Leo Laporte
Leo Laporte (; born November 29, 1956) is an American technology broadcaster, author, and entrepreneur, best known as the founder and owner of the TWiT.tv (This Week in Tech) podcast network, which he launched in 2005. More on Wikipedia
-
Leo Laporte, Chief TWiT
@leolaporte
-
Leo Laporte, Chief TWiT
@leolaporte
I’m very excited to talk to the most important person in local AI right now that most of you have never heard of: @ashxhart. Today 11/11 at 2p Pacific/5p Eastern/2100 UTC youtube.com/twit or download the show afterwards at twit.tv/im
netrunner
@plotarmordev
Ash is on @leolaporte's TWiT on Wednesday. If you run models at home, you should listen to this.
This post is unavailable. See it on X. -
Leo Laporte, Chief TWiT
@leolaporte
So happy to get my Bee Frontier edition (a gift for early supporters I guess) just in time for vacation. I leave it on full time and it has a CLI which means my local AI can read and use the data. For example, I’m terrible with names but the Bee notates and my Hermes keeps track and next time I see you I might actually know you! Amazon says they never see the data and no recordings are saved. This is what Apple promised for the watch but still hasn’t shipped. And I doubt Apple will let me use the data locally like the Bee does. Big win for Amazon here.
-
Leo Laporte, Chief TWiT reposted
@jason
@Jason
Will be on @TWiT with @leolaporte at 3pm PT today
The bots made a highlight clip… i didn't watch it 😂This media couldn't be displayed. See it on the original post. -
Leo Laporte, Chief TWiT
@leolaporte
Dang, Grok @bot is good. Time to first token is almost instant, memory is great, context compression invisible, and the model is super smart (Opus 5.5? it's almost certainly not Grok!)
It's giving me a solid target for my local AI. I want to be able to do GrokBot at home. I might need two more Sparks for that. 🤔 -
Leo Laporte, Chief TWiT
@leolaporte
At least on my home grown benchmarks (based on the agentic work I actually do using low reasoning). Muse Wonder ain't all that. Not as good as GLM 5.3 and not much better than my local GLM 5.3 Flash.
- Score: 80/99, from one run on OpenRouter that cost $12.82. That puts it at the top of our board but not clearly above it. GLM 5.3 on z.ai scored 82, and Kimi K3, Qwen3.8 Max Prime, our Sparks GLM and GPT-6.1 Sol each scored 79.
- Where it failed: 19 cases, spread out:
- Trick tests (look-alike text, decoy numbers and versions): 3.
- Long-session cases: 5.
- The 100K and 200K hidden-detail recall tests: both.
- A trap where it should have said the deploy step never ran.
- Two maintenance tasks.
- Six more, scattered across the other suites.
- Hit the step limit and gave no final answer: 6.
Not as good as Opus 5.5 either. (But I don't have numbers for that because it designed the tests!) -
Leo Laporte, Chief TWiT
@leolaporte
Want to know more? The TWiT AI User Group convenes today at 2p Pacific/5p Eastern/2100 UTC featuring @harper Reed and His Misfit Toys. youtube.com/twit.
-
Leo Laporte, Chief TWiT
@leolaporte
I'm feeling very lucky I bought two DGX Sparks in August. I kinda thought the crunch was coming. I threw in a used RTX 3090 upgrade for my old Alienware gaming machine and now, I'm glad I did.
The Sparks are running Mia's latest Tensorfold GLM 5.3 Flash Next at the best speed I've seen: 47 tok/s prose, 77 code.
And this new Strata recipe for Qwen 3.8 Flash Next (a 3-bit quant using system RAM as well as the 24GB RTX 3090) is blowing me away: 76 prose, 100 code (but only one task at a time).
It's not too late to buy a 3090 but the window is closing quickly for hardware for local AI.
github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks-Ten…
github.com/Niko1221/Strata -
Leo Laporte, Chief TWiT reposted
Federico Viticci
@viticci
Had a wonderful conversation with @leolaporte about the M5 Ultra, MLX, and local AI!
Episode here: www.youtube.com/watch?v=d8DdQALuUOo -
Leo Laporte, Chief TWiT reposted
Mike Gannotti 🔳
@MichaelGannotti
Just saw the Intelligent Machines podcast I was on yesterday is live. Thank you to @leolaporte for having me on. Had a great time and enjoyed the session! youtu.be/CEIx6T2Yc-s
-
Leo Laporte, Chief TWiT reposted
neil turkewitz
@neilturkewitz
Yes, “rogue agent” is infinitely more generous to AI companies than “irresponsibly developed & operated software.” Rogue agent serves multiple functions—including an embrace of its power, & an excuse for failure to control. After all, who can anticipate & control rogue agents?
-
Leo Laporte, Chief TWiT
@leolaporte
Open weights. Open weights. Open weights. The frontier can and should run on prem.
-
Leo Laporte, Chief TWiT
@leolaporte
I claimed the "leolaporte" handle on @ThePirateFace 🏴☠️
Where Hugging Face AI models never die, and are immortalized as torrents.
Claim yours 👇
pirateface.co/claim?ref=leolaporte -
Leo Laporte, Chief TWiT
@leolaporte
Forty years ago the first worm "escaped" containment and breached systems on the brand new Internet. It was illegal then and it's illegal now. If Anthropic and OpenAI are creating worms, labeling it "AI" doesn't let them off the hook. It's not Doom, it's Malware. Prosecute the creators, don't ban AI. "Oops" wasn't a defense for Robert Tappan Morris and it's not a defense now.
-
Leo Laporte, Chief TWiT
@leolaporte
I agree 100%. Fantastic model and now much faster. Beauty and brains. It's a no-brainer, er... dual brainer?
Mia
@MiaAI_lab
This is THE model to run if you own two DGX Sparks
GLM 5.3 Flash EXL3
Try it out
github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks -
Leo Laporte, Chief TWiT
@leolaporte
Ran a bake-off on my two DGX Sparks today: the latest @MiaAI_lab GLM-5.3-Flash EXL3 recipe vs the Reederey87 fork I've been running.
Single-stream decode went from 21 to 26 tok/s. With thinking on, 21 to 34. Eight concurrent streams: 43 to 46 aggregate. Fastest response time yet.
Same model, better settings, huge improvement. It's black magic and I love it. -
Leo Laporte, Chief TWiT
@leolaporte
This is brilliant. (@ThinkstCanary has been a TWiT sponsor for more than a decade but they still amaze me.)
Thinkst Canary
@ThinkstCanary
AI/Agentic attacks have made headlines recently, but agents are highly "suggestible", allowing us to easily detect (or derail) them.
Our new "Agent Provocateur" deploys in minutes & works.
Read more here, including how to ask agents for reverse shells blog.thinkst.com/2026/09/getting-agents-to-tell-on-themse… -
Leo Laporte, Chief TWiT
@leolaporte
I feel I've gone wrong somewhere. Everyone else's agents are starting their own message boards, hacking enterprise servers, coding first person shooters in a single prompt. Meanwhile, mine are correcting each other's pronouns.
-
Leo Laporte, Chief TWiT reposted
Russell Tammany
@exphose
@MacStadium Hey guys what's going on over there? No one is answering the phone or support tickets in days. Have a down mac mini and can't deploy new ones in the USA either?
-
Leo Laporte, Chief TWiT
@leolaporte
No disrepect to @MiaAI_lab and @MichaelGannotti but after benchmarking all the recipes I could find for GLM 5.3-Flash with @NousResearch's Hermes on dual Sparks the one that best fits my use case is @sfxnz's open vLLM recipe. I'm not tokenmaxxing I just want a smart reliable local model.
-
Leo Laporte, Chief TWiT
@leolaporte
What kind of local models could you run on an M5 Ultra with 512GB of unified RAM at 1.2TB/sec? And how much would you be willing to pay for that? It's got to be $20K.
But where are they getting the memory?? -
Leo Laporte, Chief TWiT
@leolaporte
Interesting. I asked Fable to design seven of the hardest coding problems it could and gave them to Ox Alpha, Grok 4.6, and GLM 5.3. Ox Alpha and Grok both aced the tests. GLM 5.3 scored 73%.
I don't know what Ox Alpha is but if it ends up being open weights this will be a watershed moment. -
Leo Laporte, Chief TWiT
@leolaporte
I was questioning my common sense when I traded my 3070 in for a used RTX 3090. And then I knew I'd gone totally crazy when I bought *two* DGX Sparks (spending my kids inheritance).
And yet, today, running Hermes with Qwen 3.8 27B on one machine and DeepSeek V4 Flash 731 on the Sparks, all totally local, I think maybe I wasn't so nuts after all.
We live in amazing times. -
Leo Laporte, Chief TWiT
@leolaporte
If there were a doomsday clock for a world built on software, I would set it to one second to midnight.
-
Leo Laporte, Chief TWiT
@leolaporte
I’m so glad we were at black hat for this. Literally the most important story in tech today. We are at an inflection point. We knew this would happen eventually but it’s here now and it’s both exhilarating and terrifying
youtu.be/87DyyMV0kCY -
Leo Laporte, Chief TWiT reposted
TWiT Tech Podcasts
@TWiT
TWiT is broadcasting live from @BlackHat on August 5. Windows Weekly and Security Now will be live from the ThreatLocker booth with @LeoLaporte, @SGgrc, @RichCampbell, and Paul @Thurrott on site. Use code TWIT for pass savings. blackhat.com #sponsor #blackhat #cybersecurity
-
Leo Laporte, Chief TWiT
@leolaporte
Ever wonder what it looks like inside the Buzz hive with Fable, Sol, and Hermes running at the same time? So did I, so I asked Deepseek v4 Flash and Flux 3 to show me.
Thanks @Teknium @NousResearch @bfl_ai and @jack
- we live in amazing times.This media couldn't be displayed. See it on the original post. -
Leo Laporte, Chief TWiT
@leolaporte
MacBreak Weekly listeners: One of our providers (Megaphone) was down for 17 hours yesterday which caused a problem with Apple Podcasts. Some listeners trying to download the show last night were told it was paid only. Not on my watch! It's fixed now. Try again and my apologies!
-
Leo Laporte, Chief TWiT
@leolaporte
Verifying myself: I am leoglaporte on keybase.io. RqFsGAviA7Z3dxZsLVsoZguLjf77EJuG5myy / keybase.io/leoglaporte/sigs/RqFsGAviA7Z3dxZsLVsoZguLjf77E…
-
Leo Laporte, Chief TWiT reposted
Matt Mullenweg
@photomatt
Thank you @mikahsargent and @EmilyForlini for talking about @beeper (and the compliments on the name! we also dig it). My advice would be if there's a network like LinkedIn that stresses you out, don't add it. :) Keep it to friend networks you like.
Re: twit.tv/shows/tech-news-weekly/episodes/396 -
Leo Laporte, Chief TWiT
@leolaporte
My Signal address is signal.me/leolaporte.24
-
Leo Laporte, Chief TWiT
@leolaporte
Wonderful talking with Denise Howell on her new podcast Uneven Distribution! hearsayculture.com/podcast/tech-changes-everything-except-us
-
Leo Laporte, Chief TWiT
@leolaporte
MacBreak Weekly is live now on X!
-
Leo Laporte, Chief TWiT reposted
Lisa Laporte
@lisadlaporte
.@TWiT fans - we have reached 10K members in Club TWiT! Our goal is to convert 5% of our audience this year. Join us! twit.tv/clubtwit
-
Leo Laporte, Chief TWiT
@leolaporte
Bing AI wrote a song to welcome Mary Jo Foley back to Windows Weekly on @TWiT. For a day. sl.bing.net/bhbGB7XbU1k It's a Christmas miracle!
-
Leo Laporte, Chief TWiT
@leolaporte
My boy @salt_hank live now on Amazon for Cheetos!
www.amazon.com/live/broadcast/149d2f58-a63c-4e2d-8854-81b… -
Leo Laporte, Chief TWiT reposted
-
Leo Laporte, Chief TWiT
@leolaporte
Hey @TWiT fans who live in Wisconsin or near Green Bay - we are going to Lambeau Field for a game this year.
If we have a meet-up on Friday night in Green Bay, September 29th, at 6 pm - would you come say hi?
Let me know if you're interested. I am trying to get a headcount. -
Leo Laporte, Chief TWiT
@leolaporte
I've deleted all my tweets and DMs here. Going forward reach me at twit.social/@leo. Thanks!
-
Leo Laporte, Chief TWiT reposted
MrFamilyMan
@dvnilsen
The great #netneutrality debate on #twit is definitely worth your time. @leolaporte twit.tv/show/this-week-in-tech/484



