tempoponet

My take as well. The dual-3090 and dual-spark configs have an outsized voice in this conversation. I'd add pipeline parallelism vs tensor parallelism and in the case of TP setting your lanes and speeds per card.

I might have said the same thing in 2023, but any coding CLI installed on Linux can rip through this work and keep it updated.

Most popular would be Qwen 9b or Qwen 35b with experts offloaded to CPU. Even these aren't great, but they can be run at higher quants on lower hardware. We'll see what Qwen releases version 4 in the coming weeks.

Influencers have latched onto the pitch that everyday people can run frontier models on an 8gb GPU while sticking it to the labs. There's a large audience of people who haven't had the hardware to test larger quants to see the difference.

They would be better served with smaller models that can reliably call tools and generate structured outputs without looping or totally hallucinating. These projects exist, but aren't getting amplified.

Sure, like in the case of a rear/front plate switch on a stolen car, Flock has every bit of information to identify that this is the case before deploying an officer to perform a pit maneuver on an innocent citizen. But they have no incentive to build that, and the author makes no allowances for this kind of self-imposed accountability. The entire tone is "bad things happen, what can ya do?"

The article takes a hand-wavy attitude towards abuse and negligence, with little attempt to counter-balance it with transparent accountability. If we can't address that in good faith, I see no reason to give the other 98% of the article weight.

lazypi has a sensible default collection of plugins and configs to get started

While I know it's not apples to apples, the target comparison right now is 2x DGX Sparks. Similar price, 256gb. The conversation has focused on memory bandwidth vs. compute in agentic loops, so for most people the raw numbers will mean less than the "time per task" in coding benchmarks.

This is a great article and bodes well for the M5, but we should expect more like this comparing to other platforms before we truly understand where it fits.

A lot of people just want to play "Smash Bros" on their iPad or get native ulatrawide support on "Majora's Mask".

This might be a product vs. eng split, but "whole point of the endeavor" isn't the same for everyone.

Can you share what tools and processes you're using to do this?

I've been using Pi to build custom extensions and wrapping workflows in shell processes to make it more deterministic and enforce certain validations, all guided by Fable. This isn't production work, though, just playing llm factorio at home.

It's exciting, but that's $9k for a model that's in the ballpark of last year's frontier. For many people that's like 10 years of up-front cost in a landscape that's getting better and cheaper every week.

but compared to the $500k+ racks running the current frontier, it does give perspective.

There's great VST plugins like "Sketch Cassette 2" that will do hiss, flutter, all of that, along with some great presets. That one's less about creating authentic sound profiles, more like creating lo-fi tracks and instruments for music production.

If llama 3 70b were available for $400 today, people would make it work. They would have 4 of them working side by side or 1 of them working with a GPU on another model.

To me, it's like imagining if Sonnet 3 was burned into an ASIC 8 years ago and then never changed. It would still be revolutionary, and today we would have an entire ecosystem of tools and services built around it, likely surpassing some of our current workflows.

The frontier is a different beast, but it would likely mean competing on price.

The barrier to entry for tinkering on Linux is so much lower now for anyone with a $20 Claude/Codex plan. What might have been an hours/days of fruitless head-scratching can now be done in minutes while you grab a coffee.

While I haven't tried this yet, I do have a Bazzite box that would be a good fit. I wouldn't treat the rough edges as a deterrent the way we have in the past.

In Bazzite I've had innumerable issues with waking from sleep and display driver crashes, which means a 50 ft walk to press the reset button.

If you're cool with that, this appliance isn't for you.

I bought an irresponsible pile of homelab equipment in 24/25. Hard drives, SSD, memory, GPUs.

I feel bad for people locked out right now, since it's become more interesting and important than ever.

At the time it seemed wasteful, but I'm happy to report that I'm putting it all to use now.

To go further down this pipe dream - Anthropic / OpenAI would buy them all and still price out the consumer. There's no end-run in this scenario.

I can actually use and enjoy Linux. The "year of the desktop" never came for me, but instead I got the "year of the cli".

For 20 years I've used Linux in one form or another, but I've felt like I was kneecapped for the most basic things. Just trying to plug in an external drive or a second display meant hours of stack overflow and pasting commands I didn't understand.

Now I'm using several Linux machines for Steam, NAS, local LLM, development, and what used to derail a weekend project now amounts to a coffee break while Claude figures it out.

It's fine for dense models where you need them in VRAM, less so for MoE where you're offloading layers to ram. But 32/32 is pretty good for both in the popular ~30b range right now.

- You can send any amount of money to anyone in the world very quickly and cheaply, and nobody can stop you.

- No government can dilute it or limit its supply.

Stuff like that. Maybe that matters to you, maybe not, but BTC was created because that didn't exist. And even if you don't use it, you're living in a world where financial institutions have to live alongside an alternative that does these things, for whatever that's worth.

This is where a lot of the crypto scammers on x have moved, selling ai/bots to do this. It seems like the odds slip significantly after these large bets are filled.

For the huge percentage of devs using vscode, switching to Cursor was essentially adding a new color theme and a chat window. The CLI switch was far more radical.

Thank you for the thoughtful reply. I have some smaller 3080's I'm looking to place and this sounds like a good opportunity.

The memory upgrade is $1k on a Macbook Pro. The laptop is ~$5500.

Expect to pay $4k-10k

- Your RTX 6000 is closer to $10k now

- Sparks are creeping into the $4-5k range

- AMD Strix are ~3.5k

- Apple depends on chipset and memory. Sweet spot would be 128gb M3 Ultra, probably $6-8k but admittedly haven't been tracking closely. New M5 might come in the fall. You can get a new 128gb M5 Max laptop for ~5-6k today.

- a 4x3090 rig would take $5-6k

Every platform has tradeoffs, but it's mostly ecosystem, memory bandwidth, and power consumption. They're all slow. The best option is likely to rent hardware on Runpod. The RIO on self-hosting is very low unless you have a specific need or you're ok treating it as a hobby.

Why this entire tool chain instead of building within something like pi code?

I've been exploring this area and a project like https://github.com/itayinbarr/little-coder (not my work) lets me mix and match with my current setup or any plugins built for pi.

It's older, but Hermes Pro 2 (same lab as Hermes agent) is a fine-tune of Mistral 7b for tool calling and structured outputs.

This isn't for agentic loops, though. This is for turning simple requests into API calls.

Similarly, the agentic coding success stories are from orgs that had all of these things out of the gate.

It reads like the inventors of Claude can't get Claude to apply a "human in the loop" workflow.

You can fine-tune a model, but there are also smaller models fine-tuned for specific work like structured output and tool calling. You can build automated workflows that are largely deterministic and only slot in these models where you specifically need an LLM to do a bit of inference. If frontier models are a sledgehammer, this approach is the scalpel.

A common example would be that people are moving tasks from their OpenClaw setup off of expensive Anthropic APIs onto cheaper models for simple tasks like tagging emails, summarizing articles, etc.

Combined with memory systems, internal APIs, or just good documentation, a lot of tasks don't actually require much compute.