Is the path hidden from you? AI | Local models | Freedom
Joined June 2026
- Tweets1.1K
- Following209
- Followers794
- Likes1K
Pinned Tweet
Do not pay $12,000 for 2 x DGX Sparks.
For the same price you can buy 4 x 170HXs and a computer to run them in, for FAR better performance.
Should I post complete builds at various price points?
I’ll buy DGX Sparks at their fair value.
RTX Pro 6000 1.8TB/s at $14,000.
DGX Spark 273GB/s… lets call it $2,500.
:)
The future I dream of has, in so many ways, been shaped by this man, his character, and the characters he played.
Compassion, curiosity, courage. The world would be a better place if we lived up to his example.
Rest in peace.
Seems I wasn't the only one who cancelled my subscription when this was announced, and OAI panicked. This backflip is all well and good, but doesn't change the fact that the models still suck. I'll consider resubbing when they've improved.
Day 25/
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.
Curious that the people saying we are headed towards a hyper abundant future also happen to be wealthy (or striving for wealth) and don't seem to be in a hurry to part from said wealth.
The figure I value above everything else for local inference is memory bandwidth.
This is why I believe that buying a DGX Spark is a suboptimal decision in the longer term as compared with buying multiple GPUs.
Over time, the performance improvements from larger models will trickle down to smaller ones.
In that case, you will be more constrained by your hardware's potential speed (set by memory bandwidth) than by memory size/capacity.
DGX Sparks pick the 'wrong' side of this tradeoff.
Soon: a tiny model that destroys the systems it infects, spreads across the internet, and fights back when you try to remove it.
2 x 16GB SXM2 v100s cost only $400.
That is insane value relative to how everything else is priced at the moment.
If I didn't have any hardware for local inference I would jump on this right now.
Morrowmake retweeted
Be honest: has AI increased or decreased your understanding?
I believe a simple truth: abstraction kills understanding.
People who learned computers using graphical interfaces never used a command line and had a shallower understanding of how computers work as a result. People who learned using mobile phones, which expose even less of their inner workings, knew still less.
Now we are coming to a world where you can ask your computer for anything, in any domain where computers are used, and get a passable result.
You never need to look under the surface to understand how that result was produced. Most people never will. And so most people will never learn or understand -- not just computers, but anything they ask for.
I find myself doing this increasingly out of a desire to get to the next result more quickly. But I recognise this as a failure, a habit that ultimately robs me of my faculties.
What can we do to make our AI use a form of learning and growth, not just another form of consumption?
What do you do?
Morrowmake retweeted
This space moves fast. You need to be willing to try new things and change your mind in an instant. Failing to adapt now is death.
Anthropic models crush OpenAI models and it's no contest. That is exactly the opposite of my view 3 months ago.
Be honest: has AI increased or decreased your understanding?
I believe a simple truth: abstraction kills understanding.
People who learned computers using graphical interfaces never used a command line and had a shallower understanding of how computers work as a result. People who learned using mobile phones, which expose even less of their inner workings, knew still less.
Now we are coming to a world where you can ask your computer for anything, in any domain where computers are used, and get a passable result.
You never need to look under the surface to understand how that result was produced. Most people never will. And so most people will never learn or understand -- not just computers, but anything they ask for.
I find myself doing this increasingly out of a desire to get to the next result more quickly. But I recognise this as a failure, a habit that ultimately robs me of my faculties.
What can we do to make our AI use a form of learning and growth, not just another form of consumption?
What do you do?
Idle compute is a wasted resource. That is the rule I have put in my AGENTS.md file, and you should too.
All of my agents are required to check:
🔸 How many CPU cores are free?
🔸 How much RAM is free?
🔸 How many GPUs are free?
Then reserve the amount they need for their work. This helps with 2 different problems.
First, it forces agents to utilise available resources. No more single threaded applications on a 112 core CPU. No more idle GPUs when they could be doing valuable work.
Second, it helps to manage contention. Agents see when a resource is fully occupied, and add their task + resource reservations to the queue.
This has made parallelising work using an orchestrator agent way more efficient.
Anthropic IPO and Fable 5.5 release will likely mark the top for subscription value.
From here on, the pressure to turn a profit will dominate and enshittification will follow.
Demand for local AI is going higher from here.
This space moves fast. You need to be willing to try new things and change your mind in an instant. Failing to adapt now is death.
Anthropic models crush OpenAI models and it's no contest. That is exactly the opposite of my view 3 months ago.
