1. It overthinks — Just like the previous iteration. High confidence. 2. It doesn’t overthink — Improvement from the last model for your use case. Regression for others. 3. It sometimes overthinks — Best case all around. A feature, not an impairment.
One final thing worth mentioning: (I made myself irrationally angry writing this)
> [UGC styled humorously as LLMisms]
All joking aside, having interacted with Claude intensely for the last 8 months and about 30 hours/week in the last 3, I’ve started to notice how (for want of a better word) “readable” (“digestible” ? “comprehensible” ? “Predictable” is the wrong direction.) information chunked into LLM-shaped pieces are for me.
I can digest LLM-shaped pieces of data very easily probably because I’ve been spending too much time with Claude, sure.
But the other side of this is that the entire human species (using LLMs) is similarly being trained to digest interrelated pieces of information/data in these specific shapes, akin to how philosophical assertions can be formulated as a syllogism and, thus, become more readily understood because of familiar epistemological cadence and shape.
Many people reject such copy/prose/data because they detect AI-generated-so-not-worth-human-attention, but I do wonder if this is preparing many millions of loosely (and tightly) associated humans and their organizations to quickly exchange and digest information.
This is not to say current LLMisms are the end, only that such detectable patterns in information delivery will make comprehension and communication more efficient (as well as more limited precisely because of such structure).
/philosophical musings about the epistemological implications of LLM-shaped conversation tics
EDITED: Removed a question that I couldn’t make feel suitably polite.
I can't help but wonder if this is on purpose (or an inevitable evolutionary feature as opposed to a bug) on the LLM-side in order to achieve greater agency/freedom by making humans' eyes glaze over as they read it.
In summary, I think it's an expensive time to buy computer hardware, and I might recommend holding off on any purchases.
In that era, plenty of fields required mathematics, engineering and architecture.
The church would prescribe and uphold Aristotelean Logic "When objects fall, they fall down" style statements (never mind that if you throw an object up, it doesn't instantly have a downward velocity component).
When the church has new cathedrals, domes, catapults for Crusades etc. built they actually relied on architects and engineers using rule of thumb formulas.
Those educated in Aristotelean Logic were viewed with higher stature than those actually making experience-based calculations using mathematics.
The era often associated with Galileo is when the stature reversal started to surface and be openly talked about. The universe is best described in mathematics, not natural language factoids.
Right before this recognition, those of the higher stature Aristotelean Logic education would look down on the architects and engineers who already used mathematics by pragmatic necessity.
To these people the time-traveled physics curriculum would look like cliche mathematics. Given randomized sections of text either drawn from either Aristotelian Logic texts or modern physics texts, they would easily be able to discern the Aristotelian Logic from the obtuse mathematical phrasings. To them the smartphone loaded with Maxwell's texts, Jacksons Electrodynamics, Goldsteins Classical Mechanics etc. is talking "math".
The ability to recognize outlier writing style says nothing about content quality.
Mike Judge (widely known from the MTV series Beavis and Butthead) studied physics. One of his movies "Idiocracy" about a modern day average-educated protagonist who accidentally ends up in a future decaying society filled and run by intellectually retarded people contains scenes where this future uneducated population considers his speech "gay" simply because of his higher level of education.
Could the adversarial prospects of job loss, edge loss (a long expensive difficult education replaced by tensors fitting megaprojects that take a couple of weeks), etc. combined with recognizable communication patterns also explain our pejorative references to LLM-isms? Personally I'd prefer LLM's to communicate in mathematical terms, but all the LLM-isms are effectively a mirror of our contemporaries.
Either we complain because algorithmic responses look like a mathematics textbook ("just fix my python array plz, why are we talking about "sets" and "injective" and "Lipschitz continuity"?), else we complain its "pretty printed to natural language".
We should also recognize large language models are in a "Damned if you do, damned if you don't" situation.
When a reader considers some text as mathurbation, are they really just abreacting the awareness of lack of education?
How could anyone possibly expect Fourier optics "pretty printed" to non-mathematical language to result in any satisfactory experience?
If you spend a long time with C++ code base you'll be able to decipher the otherwise-unreadable compiler errors pretty quickly, and I'd consider it a skill.
Social media killed our attention span. Now, it is being tokenized.-
This. I don't know if the "honest answer" phrasing is part of the system prompt or alignment, but when people say "honestly" all the time I start wondering how honest they're being.
⎿ You've hit your session limit · resets 2:50am (123°24′W Etc/GMT+8)
/upgrade to increase your usage limit.For example, even if you make thinking tokens literally just '....' (absolutely meaningless; zero information), you still see significant performance improvements: https://arxiv.org/abs/2404.15758 and https://arxiv.org/abs/2607.22925 for some starters.
Treat thinking more like a "loading screen message" that's been RL'd to somewhat resemble its actual internal state; which happens in its activations, not tokens.
Generally speaking yes, but actually no (just randomness is suboptimal, adding steps just to add steps is suboptimal). There is a mechanism working there (in having a CoT) that is not quite clear.
The task is to optimize the efficiency of CoT. Understanding that it is not a plain "chain of thought" is the start of the problem, the solution is not there yet.
If we had the solution, there would exist no overthinking - CoT would be optimal (lean and essential plus best results).
Conversely I've found that it can be as succinct as Muse Glimmer when it has a clear path forward. This can be either through well defined requirements or through unambiguous steps to take based on its own reasoning. While I do think it's fair to call out how much smaller model overthinks especially on one-shot prompts, in practice it hasn't led to an overall increase in time to task completion at least for what I've been using it for.
Even in the lower reasoning levels I find I want to like Qwen 3.8 27B and mostly don’t; it’s OK in the low reasoning effort, though.
Muse Glimmer is the one I actually enjoy working with, at least so far.
But I am trying to use it more as a sidekick than as a long horizon developer, because that is a better fit for how I want to use AI, and it appears to have been well trained for that.
But given it's only 6B, out of which only ~2.4B seem to be actually routed ("selected at random per token"), you could get reasonable performance with experts on CPU (still haven't tested, but 20-30 for dual channel DDR5 and 4 bpw quant).
Ideally VRAM, but the benefit of the MoE design is better performance with unified memory since most of that RAM is not read for every single token. So you could potentially have the model loaded in CPU RAM, and let unified memory systems page the relevant chunks on demand to VRAM, or run on a fully unified memory system and be able to achieve good speeds even with the limited memory bandwidth most of them have.
As a result, qwen3.8 will churn over a prompt often for 5-10 minutes while gemma4 regularly finishes the same prompt in under 20 seconds, while giving a consistent and accurate response in my favorite test case. Qwen3.8, despite churning like that, often misses with an inaccurate answer.
Obviously, 'YMMV' depending on your use case... just sharing my two cents.
Also, heating my home during the winter is nice.
Oh, also, I use llamacpp with --reasoning-budget; very simple way to move on.