Jon Decker

It comes down to use case. If it's for private use with limited concurrent requests, I would take the Mac all day (I already do at home). But if you need to support a 10-20 man dev team with many concurrent requests, dedicated GPUs and even the DGX Spark will do better (we do this as my job with a Spark serving a 15 person dev team). Like everything in AI it really comes down to the workflows you will be running. Most people using local models fall under the personal use category and would be best served with Apple Silicon (better coat, less noise, less heat, less electricity usage, etc)- but it's quite limited in higher concurrent use cases where overall throughput matters.

Some guy really said Macs are bad for local LLMs. Zero heat compared to GPUs, no PC building hassle, and unified memory lets you run massive models. Plus, the power bill is cut in half—that’s hundreds of dollars saved a month. macOS is basically perfect for agent workflows, and MLX is clearly the winning format right now. The ONLY downside is prefill speed. Name a better $4k setup than the upcoming M5 Max Mac Studio (128GB RAM, 600GB/s bandwidth). You can’t. Recommending a DGX Spark instead is wild. Good luck running agents properly on 273GB/s. (And yeah, I don't recommend the Mac Mini either, bandwidth is too low). If you're going to criticize, at least do your research first.

@IrwinDSimon Take care of your shareholders you POS. You can't make an announcement stating that no more dilution will take place?? The board needs to toss you out on your ass, your company exists because of your investors, and you can't even have the courtesy of an update. Resign and let someone run the business that knows something about investor relations. #Scumbag

Thank you @LizClaman, for the opportunity to speak with you today on @ClamanCountdown.

$tlry @IrwinDSimon All $TLRY needs to do is communicate that they have no plans for dilution to help ease investor sentiment - but not a word from them. This is the most poorly run publicly traded company I've ever seen. Not a care in the world for investors #FireIrwin

How long will $TLRY keep turning a blind eye to investor concerns? Haven’t seen any effort to address the falling stock price. @tilray

@IrwinDSimon @tilray $TLRY With the latest price action in Tilray shares I think it is time for a change at CEO This is becoming sad at this point The person responsible for creating shareholder value has completely lost any shred of confidence the street has in him

@thestreamingdev I wonder how this, combined with Google's new TurboQuant would do on larger models?

I ran a 35-billion parameter AI agent on a $600 Mac mini. Specs: M4 Mac-Mini 16GB RAM The model doesn't fit in RAM. It pages from the SSD at 30 tokens/second. On NVIDIA, the same paging gives you 1.6 tok/s. Apple Silicon gives you 30. That's 18.6x faster. No cloud. No API keys. $0/month. Here's what it can do 🧵

Needs more LIDAR! This is fresh off the heels of another Waymo collision with a tractor trailer on the freeway. No idea if Waymo is at fault in either of these, but more and more it's coming to light that LIDAR isn't some magic pill that make autonomous vehicles safe - it all comes down to intelligence

@Cannabisref @POTUS @PamBondi @GOP @WhiteHouse @PressSec @RobertKennedyJr @joerogan @elonmusk @Bruce_LeVell @FmrRepMattGaetz They have lost my vote, even if they finally act tomorrow. I have zero faith in this administration at this point. Midterms can't come fast enough

I'm also in the full-range-of-motion club, though I do mine after complete exhaustion at the end of my chest workout. I haven't tried what it feels like fresh in quite a long time. However, I agree with others- this isn't the only way to see results. For example, lock out at top could be argued that you're getting a short rest in between reps. To each their own! I agree though, 50 good-form (again, not necessarily any specific definition) is quite difficult

50 pushups in a row is extermely hard to do. If you disagree, I guarantee your form and range of motion is dogshit. Chest to ground, body in a straight line, full lockout between reps. Still insist it's easy? Post a video.

@johnrushx The issue is t the code quality, it's the overall architecture. Many times you will get what you ask for, but in a way that didn't have the forethought for extending in the future. Code quality is great, overall planning and architecting has a long ways to go

Every coder saying “ai generated code is low quality slop…” Have you actually tried comparing an average(!) human written code with an average opus4.6 generated code? I’ll bet my one kidney ai beats the average humans in the most embarrassing ways

@DirtyTesLa This will be a class action suit if they rug pull. I'm also a loyal Tesla owner and investor, but this is unacceptable. Absolutely pitiful for a company such as Tesla to treat their customers this way. They deserve whatever comes their way from this

Despite the website and email from Tesla saying Order before March 31 to transfer FSD "Tesla will NOT honor FSD transfer if the initial delivery window was after March 31, 2026 when the order was placed." Tesla just loves to treat customers like crap. Unbelievable

People have the right to complain if they ordered before that policy was changed. In fact, I believe they have a case for a class action suit if Tesla follows through with this rug pull. I agree, now that the policy is in place future orders don't warrant the transfer, but for those that ordered before the change the only right thing for @Tesla to do is honor the transfer for any orders placed before that change

I know a lot of people who are complaining about this. My opinion: There’s nothing wrong w/ keeping your current car if you love FSD that much. Remember how long it took FSD to arrive on the Cybertruck? There was a reason for that. It’s a separate build and likely a pain for Tesla to manage that branch. People shouldn’t have such a strong need to transfer FSD anyway… It’s like buying TurboTax software in 2019 and expecting you can transfer it a new computer in 2027. Consider this $60k saved. The good news is that you can use that money toward a Cybercab, a Model YL (if it comes to NA), or another killer deal Tesla offers in the future. 😉

@SawyerMerritt This will turn into a class action lawsuit, mark my words. I placed an order before this change was made as did many others. Tesla has really abandoned their old stance of taking care of their loyal customers

UPDATE: Last night, Tesla changed its FSD transfer policy (pic on left). You now have to take delivery by March 31, 2026 to be eligible. For people who placed an order before this policy change, including Cybertruck dual-motor AWD order holders: • If your delivery window is on or before March 31, 2026, or it gets pushed into April due to a delay caused by Tesla, the company will still honor FSD transfer. • Tesla will NOT honor FSD transfer if the initial delivery window was after March 31, 2026 when the order was placed. • If your order window is beyond March 31, 2026 or you end up not being able to take delivery by March 31, 2026, Tesla will offer you a refund of your $250 order fee. Before, Tesla's website said: "customers who place an order for a new Tesla vehicle by March 31, 2026 may qualify for transferring FSD (Supervised)." Not anymore. While Tesla always said directly on their FSD transfer website page that the "FSD transfer program is subject to change or end at any time," I’d still like to see Tesla give all $60k Cybertruck AWD orders made before last night’s policy change the option for FSD transfer at time of delivery. Many people placed an order because they thought they would be able to transfer FSD. I’ve deleted the post I made a couple days ago (pic on right) about Cybertruck AWD FSD transfers to avoid anyone seeing that post, but not this update, and mistakenly placing an order. While I did link to the terms page in the thread, I should have added in my first post Tesla’s exact disclaimer wording from the FSD transfer page: “The FSD (Supervised) Transfer Program is subject to change or end at any time." I’m really bummed out for the people who already ordered the $60k Cybertruck and wanted to transfer FSD, even though Tesla is offering order fee refunds. Keep in mind, Tesla can always decide to change the terms on this stuff at any time. People can reach out to any Tesla sales advisor with any questions. They will confirm this info. Link to Tesla's FSD transfer website page: tesla.com/support/fsd-tr…

@sillydarket @bradmillscan +1 for ClawVault. It's actually hard to describe how much better it is than the stock memory system. Super easy to install too!

@bradmillscan clawvault :)

i’m having a silly problem with open claw. We’re making lots of decisions and then a week later something happens or a crime job runs that references a tool or decision we made previously and neither of us remembers why the decision was made. I have to go back into my telegram chat history, find where we made the decision to do something in the chat search and then manually copy & paste a blob to my guy because he does not have access to our telegram history chat log. There has to be a better way 🤦‍♂️

I'm not sure why everyone is complaing about Mac studio price. I run my local inference on M4 Pro minis. I've slowly built a cluster over time as spare money allows. The minis are great low cost options, they crazy energy efficiency and really do a decent job for their price point. The base M4 Mac Mini is a lot cheaper even if you don't need large amounts of RAM

I'm all about the raw power of my RTX Pro 6000's, but you can't overlook what Apple Silicon can do locally. At the price/power consumption/temperature/noise, Apple silicon is currently unbeaten for desktop.

I've been testing this very thing on an isolated agent. While the documentation has a lot to be desired, I'm highly impressed at how good it is once you have it set up correctly. Don't get me wrong, the built in memory system in openclaw can be quite good too, but also only if set up correctly (which does not happen by default oddly enough). However, my own tests are showing an amazing improvement in not only what is stored, but how much better recalling from the memory store is. I'm personally waiting for V3 to drop before actually migrating my main agent's memory, as it seems like the direction of the project is pretty much laid out

@sillydarket benchmarking etc, a stable version that is safe to integrate into an established openclaw setup.

what are the best / most difficult memory benchmarks out there? anyone working on any interesting memory benchmarks? what are the best memory frameworks we should put toe to toe with clawvault?

I have been testing it. I set up a separate agent for my tests. As I expected, even after pointing to every documentation link and blog article, it still struggles to get things set up correctly. I think it may finally be close to being set up correctly after about 4 hours of back and forth with the agent, and I finally figured out how to use Z.ai as the provider for the compressionLLM. Just wanted to share the feedback. Better documentation or instructions would be very helpful. I would think a majority of users will be on OpenClaw, so specific instructions to properly set up and migrate the old memory system are still needed for a smooth setup experience.

@jondecker76 V3 will handle this graciously

If you were to not text your claw for 24 hours, what would he actually get done for you? what if he didn’t need you to truly get work done, stay organized and follow goals?

@0xPeruzzi @sillydarket Unfortunately I'm not willing to .I grate without official documentation. My agent has screwed up more things than I can count- and without proper documentation, the agent's best guess is no better than anyone else's.

@jondecker76 @sillydarket I just asked mine to draft a migration plan

@sillydarket Great! Unfortunately I can't find anything the docs. I can see env variables for OpenAI, Gemini or anthropic API keys, but no mention of anything beyond that. Thanks!

@jondecker76 You can!

if after you install clawvault, you have any frustrating interaction where memory or context is the issue (from context death, to having to repeat urself, or agent not following a rule/pattern) please come yell at me so I can make it even better