I prototyped something like this for fun a long time ago. Treating s3 like a bucket of blocks seemed intuitive way build a scalable filesystem. Arguably ceph and luster are doing something similar except with a seperate metadata servers to serve the hotter content.
I think the critical thing you will need to explain is durability and loss window. Making some guarentees on failure modes would go a long way towards making me believe i can run operations on something like this.
With AI you should be able to do some exhaustive testing both for load, power loss, server loss, etc. Anxious to see the potential results
You shouldn’t expect frontier models to work on medical imaging. There is much more that goes into building a medical imaging product. First and foremost is data. Medical imaging datasets are not prevalent one the public internet at the scale necessary to have good performance on medical imaging tasks especially MRI. Also the labels are super noisy.
This is completely different than asking for general medical reasoning which is more derived from papers, public standards and textbooks.
Look into the Eigen library. They use template meta programming to chain linear algebra operations in a way that the compiler should be able to optimize memory layout and kernels for vector instructions. Might give you some ideas.
Ran the same query and there is a ton of stuff, but it looks like it's reasoning through the ambiguity of the sentence. It still gets the right answer. Moreover, if we consider the FLOPs expended to get to the answer, and compare that to opus, I think it's still a net win.
My hunch is that Opus scale models probably have shortcuts encoded into the model that handle these ambiguities cases, wheres this model has learned a program to reason through the edge case (crystalized vs fluid intelligence). Remembering that probablity (frontier) vs calculating it on the fly (vibethink)
I'm super happy with the performance, I generally run with 2 parallel slots so I only get about 128K context window. My experience with all llms is that they get more forgetful if you use the full window. (256-512K is the sweet spot for frontier models, 128k works for me with this current qwen)
I forked it to also add rotorquant. This is a specific optimization that uses clifford rotors instead of static compile time random purmutation to store the activations. Reduces space and parameter count for the storage.
I got qwen3.6:27B running on my 4090 (24GB) with ~128K context leveraging some of the recent turboquant/rotorquant memory optimizations for activations. Highly suggest going up to that. the q4_xl+rotorquant combo is pretty good.
Not to defend Elon, but he spent 40B on a company that dropped in value between the day of his bid and the day of closing by almost half (if i remember correctly). He tried to back out of it but then went forward by twitter at the time sued to enforce the transaction.
150M is small in comparison to the 20B he overpaid by. Investors were well compensated.
I've liked qwen and embeddinggemma for local search. Qwen because 32K is enough to basically fit a whole page into the context window and embeddiggemma because it's crazy efficient.
When I lived in europe, RyanAir made most of it's money in the terminal. This is why every RyanAir terminal has a maze like exit from security (ikea-esque) before you get to the actual terminal.
Fair point and I don't disagree. The more meta point I would make is that airlines are still fairly capex heavy (even if there is no point-to-point infrastructure). Each incremental new route operating during standard hours still requires 90m+ on a new airplane.
So if they tend to compete themselves into oblivion, or need to turn into banks to subsidize their product, then it might make sense that they should be regulated monopolies.
Still you're probably right, if they can turn into banks and stay profitable, then maybe that's a better market outcome overall.
Power companies are the classic example. If power companies were forced to compete, their costs + competition tend to drive them out of business. As a result most power companies are forced to operate in really tight constraints with very limited but predictable margin.
I'm not saying that this a better outcome (power companies have their problems too). I was just commenting that this issue parallels the historical solution that was applied to utility companies.
Fundamental problem: Flights don't make money. Airlines actually make all of their money through loyalty programs and credit card payments. They basically should have turned into regulated utilities long ago, but loyalty program revenue saved them.
Unless this initiative will turn into a credit card company (which nobody likes or wants to do) it won't go anywhere
Private equity will likely sell the company for parts. There is no operational improvements for cash flow that they can do.
The list is great but the explanation are clearly AI slop.
"Before SpaceX, launching rockets was costly because industry practice used expensive materials and discarded rockets after one use. Elon Musk applied first-principles thinking: What is a rocket made of? Mainly aluminum, titanium, copper, and carbon fiber. Raw material costs were a fraction of finished rocket prices. From that insight, SpaceX decided to build rockets from scratch and make them reusable."
Everything including humans are made of cheap materials but that doesn't convey the value. The AI got close to the answer with it's first sentence (re-usability) but it clearly missed the mark.
This is the fallacy. These organizations no longer have any ability to “legitimize” as trust is fundamentally eroded. Leaving will simply remove any engagement with the very people they want to influence- people that are unengaged and people that actively disagree
Credibility with who? We’re so polarized that a single binary label will shift all credibility.
Experience, success, credentials none of it matters anymore. The left thinks everything on the right is stupid and evil, the right does the same, and everyone drinks their own kool aid.
This is Part 2 of a series on agentic systems — This was especially weird given how eosteric it can be to describe how to work better with agents, but this is my shot at it.
The article walks through the coordination patterns to address types of error:
- Single-writer: one agent writes, others review read-only. Eliminates oscillation. Maps to
the single-writer principle from concurrent systems.
- Sequential planning: parallel planners cluster even across different models. Sequential
divergence acts as a covering algorithm — 3 sequential planners explore more than 5
parallel.
- Sequential vs parallel review: parallel voting catches common issues (mean quality).
Sequential review compounds scrutiny but risks scope creep. Both are useful.
- Human interview gating: open-ended questions yield ~5x more useful context than closed
ones. "REST or GraphQL?" vs "What should we know about how this API will be consumed?"
- Adversarial validation: separate environment, separate agent, explicit goal of breaking
the application. Tests the spec, not the implementation.
I think the critical thing you will need to explain is durability and loss window. Making some guarentees on failure modes would go a long way towards making me believe i can run operations on something like this.
With AI you should be able to do some exhaustive testing both for load, power loss, server loss, etc. Anxious to see the potential results