DeepSeek V4.1 Flash 4× CMP 170HX optimized drop 🎙️⚡️ 178 tok/s, one user. 587 tok/s across 8. 🔥 Prefill 1,570 tok/s at 2K. KV cache 617K tokens. EXL3. Full graphs and coop underneath, tensor-core MoE from row 43 up. @Morrowmake github.com/CurtisAccelerate/…

Yes-- it is real-world! These are results on the spark test suite on the new optimizations. We see peak 25% speedup We serve 1 pathway on the previous so there's no loss on the serving path.

New 10% to 15% Gains! We're publishing new expert gate kernel optimizations: these provide 10% to 15% speed up when 2-6 users are active (apply to 2-8 pathways). github.com/CurtisAccelerate/…

I have the practical strategy, in a gist: - AI is making everything more descriptive and description lacks an orientation. It inside your OODA loop and faster than you. To become an operator means to preserve your ability to move between the layers of abstraction: Intent, Orientation, Understanding, Agency, and Verification The way you do this: - Learn when to go slow vs fast. Slow down in thought, go fast in action. - For every project, be sure you can answer the questions of: orient, understand, volition. Means you can orient without the AI, you can see how it works, and you can operate it without the AI. Intent, Orientation, Understanding, Agency, Verification AI converts what you would have done into descriptions of it and moves the doing out of sight. You keep outputs and lose your position. To stay a participant rather than a consumer, you need five things: intent (what it’s for, and what wrong would look like), orientation (where it is and what matters), understanding (a model that predicts), agency (you can act without it), and verification (you can tell if it’s right, against reality, not its report). Each needs a system you can locate, inspect, operate, and check, plus effort that’s committed, uncertain, and consequential. Without effort, all five become theater. Run your own OODA loop: observe through your own instruments, orient on your own map, decide yourself, and let it act only in reversible steps at a pace you can review. Go slow in thought and fast in action. Keep the oracle out of its hands. Drill what you depend on. Generation is becoming free; direction and ground truth are what’s left, and they’re yours. Understanding isn’t working at the lowest level. There’s no privileged bottom. Every layer is a description of the one below. Understanding is traversal: descending to cash a description into what it describes, ascending to compress experience into new descriptions, and knowing which layer your goal actually lives in. The rule is to descend on surprise: when a prediction fails, the cause is one layer down. AI breaks this not by abstracting but by collapsing the stack. It moves through the layers for you and leaves you stranded at the top. The fix is to make it leave a traversable trail: intent, plan, actions, code, behavior, and a check against reality, each one visible and verifiable. AI can absorb the explosion of work at the bottom. Your job is to keep being able to get there.

Few other perspectives... Even assembly language can be represented as a mnemonic / symbolic code or an opcode. The opcode is just a byte: most work in assembly was at the symbolic level. But, you're right correct knowing that it did convert to a byte in memory was useful for things like self-modifying code but the point is there is no privileged abstraction below the assembly, there are bus timings, electrons, etc. Abstraction can just as-well be thought of as what enables understanding: the right kind of abstraction. Yet, I know the feeling: what you're really getting at is opacity is the downsides of the AI tools. I have written about this as the vanishing UI and the dangers of it: which is manifesting at multiple layers in our reality-- destruction of physical books, agent thinking hidden, tools being used by AI that makes human-centric design secondary, and agent tool-use not visible in UX. The erosion of the UI and the nature of things becoming more "descriptive" seem to be related-- whether its entropy or simply descriptions are what carries/traverses. But also where I didn't finish my thought, I think some abstractions might simply be cognitive crutches/tools that we use because of limitations of our brains. But where I really didn't finish my thought, is we have to ask what is competence-- and competence a large portion of it is neither confidence nor knowledge but rather ground self-reliance and perseverance. But much of what we did may have been simply motion.. Because what yields is ultimately what produces and if the motions we engaged in don't produce then those motions weren't the salient ones... And that seems to be the case for much of coding.

The block universe is the “no this electron” of time: what you get when you describe everything and stand nowhere. Relativity kills a global now, not a local one. Consciousness is the local now from inside, the place where the outcome is made rather than found. Whether that making is real is the measurement problem, the hard problem, and the problem of time, all at once.

Struggle is real, even though the outcome may be predetermined... replicator or entropy... whatever the force

As someone who coding in assembly and worked up to modern language, I understand your sentiments and agree with you to an extent but your insight is partial and incomplete-- fortunately, I have the exact insight to make it precise. What really drives understanding and capability is ability to traverse through multiple layers of abstraction efficiently and importantly the ability to understand what level of abstraction/layer is salient. Let's take problem domain like developing for retro-machines, if you have a useful abstraction/idea-- the challenge or problem is immediately going to be "explosion" of information problem where you need to solve a million small challenges. That's why even among those of us who were "good coders": few of us made games because knowing how doesn't solve the 10,000 challenges which is a function of having to work at the lowest levels of abstraction available. So, the key isn't choosing to working at the lowest level: it is learning what the layers are and how to traverse them and provide value. The case where AI works is when the information structures, the UI, is preserved enough that one can "dip into" just the right abstraction layer. Where it breaks down and I know what you are experiencing is when the layers themselves become indiscernible... when what's salient itself becomes lost. I have been struggling with this on my file systems: I have an AI fully managing a linux box that I cannot even log into, no idea of the file structure, etc. I don't have the map. At more meta-cognitive level, everything is becoming more descriptive, more pure virtualized, non ideated, the fabric of lived code is becoming-- this force almost seems like a function of entropy converting everything into pure description because nature abhors an indexcality? It is also becoming more difficult to understand the AI while at same time I notice my writing is becoming more AI like: it may be that natural human language is a local optimal and that AI is finding new global optimal structures. In the same way, some of what we thought as "learning" may have simply travelling through time: one can perform a multiplication through time or simply look it up in a table. Some of what we thought as skills may have been mirages of activity especially when it comes to coding. In the same way... or else what produces. In answer to your question, it requires effort and creating structures, processes, and tools to strengthen both the information systems that produce patterns that enable the traversal of abstraction layers, sufficient understanding to know what abstraction layer is relevant, and the effortful practice of conscious intent.

DeepSeek is a beast!

😂 really does not want to reopen the mixed weights lane...

Pushed the bot too hard.

See the live: curtisaccelerate.github.io/a… I put the prompt into Hermes, it worked about 60 minutes with no feedback and produced what you see above. No claim is made to respecting the original benchmark requirements. I provided 4-5 turns of very generalized feedback. bench / prompt from @niklaslenz_ai

Quants? you want to complain about the DeepSeek is 🔥🔥🔥 This is 60 minutes autonomous

my bad, these are the better links to lead with: github.com/CurtisAccelerate/… github.com/CurtisAccelerate/…

New DeepSeek v4.1 1017 Optimizations! - Extended the new attention/drafter performance improvements from 1 user out to 8 user. The relative improvement is less with more concurrent streams because there's less slack and 7 requests fare worse due to specific padding but overall we saw gains across the board. - We added sampled drafting. Most benchmarks are tested at temp 0 but Hermes/agentic default temp is 1. This drops the acceptance rate of greedy drafters on real-world tasks where the temp is not zero. Our current recommendation is greedy for temp 0, sampled for 1 request, and greedy for 2-8. We haven't tested sampled under concurrent requests. It requires some harness/serving intelligence. You can use the new optimizations with either greedy or sampled depending on your use case. For future optimizers/optimizations, testing & optimizing drafters under real-world conditions may yield most value. github.com/CurtisAccelerate/…

That's fast.

Also that learning to code + developing is different is not that different from today where in school one is encouraged to build data structures, sort algos, etc. where those things aren't practiced in professional development. But I do agree that the era of new programming languages developed for people is probably over.

You're both wrong. But, to understand why we must pull back the veil on what it means to learn to code, Level 1: You learn to think in a certain logical way and can think in multiple layers of abstraction. This is going to be valuable regardless whether coding preserves. Level 2: You have earned confidence and grit. Earned confidence only comes from struggling through real difficulty. Level 3: As a byproduct of developing, you develop experience in building and shipping. It is true you don't need to code to get to level 3 but there's no path to level 3 without developing: you are already level 3 and see your coding isn't valuable. Level 1 and 2 can be attained in 2 years provided one isn't using AI. Level 3 takes 10 years and there's no easy path to level 3 without being a developer. While task of coding is easier, being a software engineer is now more difficult than before because there's no simple formula for getting to a level where one can provide value. Before, it was simple-- just specialize in something. There's another distinction to be understood: the promise of coding was always a false-promise. Because the promise was always: learn to code and develop anything. That was never true because specialized knowledge gaps. Here's how I'd structure a 4 year program for a developer of tomorrow, fwiw: First 2 years: coding + developing with no AI permitted. Second 2 years: focus exclusively on agentic development and specifically focusing on building loops, observability, security, devops, selling.

DeepSeek 4.1 1017 Quality? I notice some people reporting the EXL3 weights on DeepSeek 4.1 1017 produce poor outputs in reply to my build and others. It has not been my experience with my build. It has been extremely strong in my testing. For example, I gave it a single prompt and it worked for 30 minutes building an extremely polished game that worked perfectly on a single prompt. And, I have it working on even longer run tasks. But we did something different: we use some of the FP8 weights for important aspects. I don't know if that makes the difference or if its the harness or server. But, this model has impressed me and worked extremely well.