Bryan

@_Adrian Not an OpenClaw guy I guess? I'm still experimenting myself.

M3 Ultra Mac Studio 512GB . Was running Llama 70B but switched to Gemma 4 31B. Stack: Ollama + Open WebUI , running Q8_0 It’s really good but not current frontier grade, not even close. My codebase already had it fumbling. I’d say it’s just below GPT 3.5. We’re getting close.

Gemma 4 E4B on device is cool. Gemma 4 31B, on an M3 Ultra Mac Studio 512GB is much better than, say Llama 70B, but NO WHERE NEAR Opus 4.6 or Gemini 3.1 Pro. These “AGI” claims are (obviously) pure engagement farming.

@tkarrdeiz @AlexFinn I have 128GB. Currently, I use Claude Haiku and Sonnet when needed (API).

@TexasBryan @AlexFinn If you have less than 32GB RAM, stop cramming mediocre local models into it. Use your RAM for VMs and agents. Offload inference to MiniMax M2.7 cloud at $8/mo. More intelligence, more headroom, and you're not babysitting a 24/35 model when a 33/35 is eight bucks.

Do you even understand what this means? An open source model just released that is: • Outperforms models 20x its size • Can run on a base model Mac Mini • Is AMERICAN 🇺🇸 If you have a base model Mac Mini you can have unlimited super intelligence on your desk. For free. Sonnet 4.5 was released 5 months ago In 5 months that level of intelligence went from frontier to free on your desk And not only that, can run on any basically any computer out there If you have even a remotely modern computer, do the following immediately: 1. Download LM Studio 2. Go to your OpenClaw and ask which of these new Gemma 4 models is best for your hardware 3. Have it walk you through downloading and loading it 4. Build apps with it knowing you are using your own personal, private super intelligence on your desk The people denying this is the future are so beyond lost.

@tkarrdeiz @AlexFinn Is there an MLX version of Gemma 4 to run on LM Studio? I just learned yesterday that the MLX versions run faster on Apple silicon. Thanks!

@TexasBryan @AlexFinn Qwen 3.5 27B is dense — all 27B params fire every token. That's why it's slow. Try Gemma 4 26B-A4B (Q4, 18GB). It's MoE — only 4B active per token. 42 tok/s vs crawling. For basic task management you don't need 27B thinking on every prompt.

@tkarrdeiz @AlexFinn if you had to pick 1? I'm really just looking for a good local agent that can manage tasks, no real heavy lifting at this point. I just installed qwen3.5 27B. Seems smart enough but responses take a good 30 seconds or more even for simple questions.

@TexasBryan @AlexFinn Swapped Qwen 2.5 32B for Gemma 4 26B on our Mac Studio. Same eval score, 4x the speed, and it catches structured data Qwen misses. Qwen still writes better client emails though. No perfect small model yet.

@Web3Twon @AlexFinn good advice, gpt is advising me to install both and set up a hybrid system depending on if I need speed vs power

@TexasBryan @AlexFinn Honestly... Just try both The 27b is going to be smarter all around because all 27 billion parameters are active at all times. The mixture of experts model has the ability to be smarter but with only 3 billion active it will probably fall short in some areas but not speed

@Web3Twon @AlexFinn I just asked gemini again and now it says I should look at 3.5 27B or 3.5 35B A3B. I'm not a heavy coder and I value speed so I'm wondering if 35B A3B might be best for me???

@TexasBryan @AlexFinn Yeah Gemini giving you outdated info it sounds like or maybe 3.5 isn't available for mlx yet If it is def run that Qwen 3.5 27b is damn impressive

@Web3Twon @AlexFinn No, it's Qwen 2.5 32B mlx. Gemini recommended this to me based on my setup with 128GB of Ram. I did question whether I should install 3.5 instead though.

@TexasBryan @AlexFinn Since Alex didn't really answer the question I'm assuming you mean Qwen 3.5 right? 2.5 is definitely not as good as Gemma 4... If you met 3.5 it's probably pretty close but this new model from Google seems very promising

@RoaringRagnar Interesting. I just spent the last week configuring Openclaw. Now I need to try Perplexity. This stuff moves too fast 🤣

I just used Perplexity Computer for the first time. It’s a digital coworker that doesn’t just answer questions. It does actual multi-step work for you across apps, files, and the web. It‘s basically a combination between Claude Cowork and OpenClaw, packaged for ease of use. It has access to all leading AI models in the background, and orchestrates them automatically, depending on the task. I‘ve built a web app, a powerpoint presentation, and an agent that looks for specific videos, downloads them and cuts them into clips. All in under one hour. 🤯 It‘s nothing short of mind blowing what @AravSrinivas and the @perplexity_ai team have built. AI agents will change everything.

@JoshMandell6 is there anything you can think of that would drive the price of STRC down to the low $90's or worse in the near term? It seems almost risk free but there is always a catch.

What's one significant difference between STRC and MSTR? Given the float in the Common it was relatively easy to borrow and short sell massive amounts of MSTR, sending price towards NAV in the process. Some have spoken about a possible death spiral and STRC, But good luck running a big short position on something that's paying 11 1/2 percent. One should realize that a short seller holding short in that preferred has to actually pay that Div every month and it doesn't have a USD cash reserve. @phongle

@TonySeverinoCMT @BrianRoemmele you should close your bank and brokerages accounts then

@TexasBryan @BrianRoemmele No, you and this guy should face reality You can’t stop this and it’s going to evolve much faster than even Google realizes Pandora’s box is open

GOGGLE CRACKED BITCOIN! Nope. Google Quantum AI just published a new paper showing ECC-256 (Bitcoin’s curve) could theoretically be cracked with <500k physical qubits. Not a big deal for Bitcoin. - Still pure theory. No one has hundreds of thousands of error-corrected qubits. - Most BTC sits in addresses where the public key is never exposed until you spend (don’t reuse addresses). - Bitcoin can soft-fork to quantum-resistant signatures years before any real threat arrives. HODL calmly. We’ve got time. Paper here: quantumai.google/static/site-as…

@TonySeverinoCMT And to push back, do you not think quantum can also find better security solutions? I don't know the answer but if heavy encryption will be broken, then many things beyond BTC are in big trouble.

If it’s lucky, Bitcoin will have a lost decade Otherwise, it very well could simply be over Yesterday, the majority thought Bitcoin would only go up forever After this quantum reality check from Google, do you still think price just continues to go up forever? Be realistic

@TonySeverinoCMT I'm pretty sure Google said this is a future risk (2029). If BTC is at risk, so are all the banks.

@HansCNelson I just set my Claw system up last week but have only set up Telegram so far. I guess I need to connect to Discord.

I'd heard that everyone loves Telegram for their Claw/Hermes agents. I hate it. It's so bad. Going back to Discord. I'll gladly pay an extra $5/month for the file storage to use a messaging platform where agents can actually talk to each and read the darn channel history.

@ChrisCamillo I'll let you know one year from now 🤣

One of my favorite deep thought exercises during times of extreme market uncertainty and fear: What trades do you expect to regret not making this week… a year from now?

@ZeroHedge_ The PE multiple on this company that prints money is getting ridiculous. With my luck, as soon as I short it, it will slingshot back up. 🤣

$MU My biggest short just hit the jackpot. It was VERY EASY with the vault indicators and our process to see this was the end for a long time A sign of weakness that bad doesnt just recover easily. Thats a complete structure break into stage 4 Bulls money becomes bear money