DebertaV3's architecture and noising should be even better as a basis because it had a couple inductive biases (cross encoder, disentangled attention and RTD corruptions) that enabled it to have unmatched weight performance ratio on such tasks.
My gut tells me that a better approach to a calibrated 0-shot classifier than shoehorning DiffusionGemma would be starting from another gemma, T5GemmaV2. Take its encoder and do continual training on an RTD objective and a large relational synthetic data mix. Then finetuning (multi-annotator data will help calibration) on as many proper NLI datasets as possible. That still lacks the DebertaV3 disentangled attention's inductive bias, however.
Jev also has its calibrated predictions component which is important. Temperature scaling is probably the easiest first pass. But there's lots of sensible options to improve on that.
ModernBERT might be the easier, more stable starting point than T5Gemma though.
I suspect you could get interesting results, but DiffusionGemma has a lot of knowledge that may be challenging to train into the smaller models. The advantage of pulling a fully-trained diffusion model off the shelf is that it already knows all of this, has been trained as a MoE, etc.
What I think these models actually need is a structured decision thinking mode. As it stands now, the only way to think about the answers with DiffusionGemma is to diffuse a thinking block, but giving the model an auto-regressive thinking space to reason, even just lightly, could drastically improve performance.
There is a video with John Baez and Philip Gibbs explaining why this is a common misconception. It's the same family of conceptual error, where things lining up just right allow incorrect conceptual reasoning to land at a correct answer shape, that allowed John Michell to work out the Schwarzschild radius of a star in the late 1700s.
With the blackhole universe picture, causal boundaries and spacetime's evolution aren't properly accounted for once we step out of a static picture.
Incidentally, the cosmic object we are more likely to be inside of is a white-hole.
QAT, which bitnet training is a form of, helps a ton in preserving accuracy at such low bits per parameter. There are also better quantization approaches that try to preserve the most sensitive weights† but are computationally expensive and so not typically done. Another complementary option is, if the model is fast enough, we should be able to push up correctness by self-consistency voting at close to T=1. Smart/fast Zero-shot classifiers like the recent Jev could help with aggregation across answers too, extending applicability.
†Every paper I've read estimates the average information content of transformer LLMs at about 3-4 bits per parameter. Curiously, biological synapses are also estimated to be about 4-5 bits per synapse, possibly a bit lower.
the average information content of transformer LLMs at about 3-4 bits per parameter
The problem is that 4-bit block-wise quantization does not guarantee preserving 4 bits of useful information per parameter - not even on average. It simply assigns one of 16 quantization levels to each weight, with the whole block sharing the same scale/range.
How efficiently those 16 levels preserve the model’s information depends on the weight distribution, block size, range/clipping strategy, outliers, and which weights are actually important. Some weights may be represented almost exactly, while others lose much of their useful information.
A simple example is an outlier: if you choose the range to preserve a very large weight, much of the 16-level dynamic range is spent on that outlier, leaving coarse resolution for all the smaller weights in the block. So 4 bits of storage does not imply 4 bits of useful information preserved. Yes, QAT helps, but usually at the cost of learning efficiency. It takes longer to train a model to the same quality when using less precision, and sometimes we simply cannot get to the same quality level with not enough precision in the right places.
Another problem in quantization is that we don't really know which weights are sensitive - we can compute various sensitivity metrics, and some of these metrics will correlate with accuracy on some benchmarks, but not on others.
Another complementary option is, if the model is fast enough, we should be able to push up correctness by self-consistency voting at close to T=1. Smart/fast Zero-shot classifiers like the recent Jev could help with aggregation across answers too, extending applicability.
I'm not convinced by this argument - if such a method improves accuracy of a degraded quantized model, then it could in theory also help non-degraded full precision model. And if so, then we are back to square one, because this composite model will then get degraded due to quantization (baseline has improved!)
We do know one thing - increasing the size of the model usually makes it more robust to quantization. If going from 8 bits to 2 bits speeds things up by a factor of, say, 4x, then if we double the size of the model, we might still end up with an overall speedup. Finding this balance might become a hot area of research.
This is partially mitigated by the fact that all the formats that quantize to 4 bits, or other such low values, partition the weights into small blocks and they also keep scale factors for each block of 4-bit weights.
This works well when the dynamic range of the weights does not vary much within a block, but it fails when closely located weights have very different magnitudes.
NVFP4 is more accurate than other 4-bit formats, because it stores more scale factors, i.e. 1 FP8 scale factor for each block of 16 4-bit weights, plus 1 FP32 scale factor for each tensor.
MXFP4 uses blocks of 32 values, and the common scale factors are only powers of two (which provides a higher dynamic range than FP8, but a coarser resolution).
The most charitable explanation I can think of for this is something like regression to the mean. When a model is first released, there'll be a subset of users who, just by chance, sample the highest quality band of the distribution that answers their query. Some of them will rush over to social media and post about how amazing a model is. Over time, those users' mental model of responses will converge but they'll perceive the model's return to typical performance as a downgrade.
This guess/explanation predicts that most users won't match what the initial social media hype claims, doesn't discount user experience as simple habituation nor does it assume companies are lying when they say there have been no changes to the model itself (quantization included).
I also think there's an aspect where initial testing is more forgiving because the more persnickety polish bits can be ignored and tests are likely to have similar structure to things that can be trained for. Meanwhile, actual specific work items are a broader unusual distribution with more stringent acceptance criteria.
Personally, I can detect a separation between Sol and Astra (but not as large as that between Opus and Fable). While they can solve most of the same problems, Astra takes less time, is less frustrating to talk to, is cleaner, notices more, spins wheels less and requires less corrections.
> I also think there's an aspect where initial testing is more forgiving because the more persnickety polish bits can be ignored and tests are likely to have similar structure to things that can be trained for. Meanwhile, actual specific work items are a broader unusual distribution with more stringent acceptance criteria.
IMO this is 90% of it (as someone who has a bit of a different interaction style and runs these things less autonomously, and hasn't generally seen the claimed regressions). Day 1: throw new stuff at it that failed badly, exciting to see something make more progress! Day n: reality sets in that it still wasn't perfect the first time.
To add onto this, if you use a shiny new model and it gives you a turd, you're not going to tweet about it ("hey guys, look what I made with Astra! Nothing!"), and even if you do nobody is going to interact with it so it does poorly in the algorithm, because it has to compete with all the people using the new model to make something that looks impressive. Then people get tired of the magic trick and the logic flips.
We've seen that some---gpt5 was considered pretty lackluster intially, in particular. Opus 4.7 and 5 vs 4.6 were also greeted with a lot more "meh" than 4.6 or Fable.
The weights + the architecture are already 100% of the code, the transformer is just a mathematical expression + helper programs whose sources are provided. The transformer itself is not even a stateful program, so a it is no more a binary than Piet or Tromp's BLC are. It's merely incomprehensible. Training isn't compilation either, since training a model is closer to program induction and the data are samples defining the solution space.
What constitutes serious work and how seriously have you tried to do serious work with them? While those trying to claim a 30B dense model can match Opus 4.6 are engaging in either beyond over-excessive over-exaggeration or performing rather routine tasks, it's disingenuous in the other direction to claim the latest open 1T models are not useful for serious work. I find those making such claims have rarely spent more than a few minutes on halfhearted attempts and often on recently obsoleted models.
Openweight models turned a corner around kimi 2.6, deepseek v4 pro/flash, hy3 and mimo 2.5 pro. Similar to how closed LLMs turned a corner around gpt 5.2 and opus 4.5.
While they remain a step behind closed frontier models, for real world tasks ranging across functional reactive programming, distributed systems, mathematical modeling, to-the-millisecond highly optimized spatial data-structures, complex compute shaders and shader effects and non-trivial systems involving parser combinators and algebraic effect systems, I can say that open models have very recently gone from useless to productive. For my work, mimo v2.5 pro is hands down better than sonnet 4.6.
Interesting thought to consider in principle but fails because gorilla brains continued to evolve too, just along a different path. They're not snapshots of ancestral species locked in time.
This sounds very like Licklider's essay on Intelligence Amplification: Man Computer Symbiosis, from 1960:
> Men will set the goals and supply the motivations, of course, at least in the early years. They will formulate hypotheses. They will ask questions. They will think of mechanisms, procedures, and models. They will remember that such-and-such a person did some possibly relevant work on a topic of interest back in 1947, or at any rate shortly after World War II, and they will have an idea in what journals it might have been published. In general, they will make approximate and fallible, but leading, contributions, and they will define criteria and serve as evaluators, judging the contributions of the equipment and guiding the general line of thought.
> In addition, men will handle the very-low-probability situations when such situations do actually arise. (In current man-machine systems, that is one of the human operator's most important functions. The sum of the probabilities of very-low-probability alternatives is often much too large to neglect. ) Men will fill in the gaps, either in the problem solution or in the computer program, when the computer has no mode or routine that is applicable in a particular circumstance.
> The information-processing equipment, for its part, will convert hypotheses into testable models and then test the models against data (which the human operator may designate roughly and identify as relevant when the computer presents them for his approval). The equipment will answer questions. It will simulate the mechanisms and models, carry out the procedures, and display the results to the operator. It will transform data, plot graphs ("cutting the cake" in whatever way the human operator specifies, or in several alternative ways if the human operator is not sure what he wants). The equipment will interpolate, extrapolate, and transform. It will convert static equations or logical statements into dynamic models so the human operator can examine their behavior. In general, it will carry out the routinizable, clerical operations that fill the intervals between decisions.
e.g. (amongst many others) Desk-Surface Display and Control: Certainly, for effective man-computer interaction, it will be necessary for the man and the computer to draw graphs and pictures and to write notes and equations to each other on the same display surface. The man should be able to present a function to the computer, in a rough but rapid fashion, by drawing a graph. The computer should read the man's writing, perhaps on the condition that it be in clear block capitals, and it should immediately post, at the location of each hand-drawn symbol, the corresponding character as interpreted and put into precise type-face.
This is essentially what any relu based neural network approximately looks like (smoother variants have replaced the original ramp function). AI, even LLMs, essentially reduce to a bunch of code like
let v0 = 0
let v1 = 0.40978399*(0.616*u + 0.291*v)
let v2 = if 0 > v1 then 0 else v1
let v3 = 0
let v4 = 0.377928*(0.261*u + 0.468*v)
let v5 = if 0 > v4 then 0 else v4...
Thats a bit far. Relu does check x>0 but thats just one non-linearity in the linear/non-linear sandwich that makes up universal function approximator theorem. Its more conplex than just x>0
Multiply-accumulate, then clamp negative values to zero. Every even-numbered variable is a weighted sum plus a bias (an affine transformation), and every odd-numbered variable is the ReLU gate (max(0, x)). Layer 2 feeds on the ReLU outputs of layer 1, and the final output is a plain linear combination of the last ReLU outputs
// inputs: u, v
// --- hidden layer 1 (3 neurons) ---
let v0 = 0.616*u + 0.291*v - 0.135
let v1 = if 0 > v0 then 0 else v0
let v2 = -0.482*u + 0.735*v + 0.044
let v3 = if 0 > v2 then 0 else v2
let v4 = 0.261*u - 0.553*v + 0.310
let v5 = if 0 > v4 then 0 else v4
// --- hidden layer 2 (2 neurons) ---
let v6 = 0.410*v1 - 0.378*v3 + 0.528*v5 + 0.091
let v7 = if 0 > v6 then 0 else v6
let v8 = -0.194*v1 + 0.617*v3 - 0.291*v5 - 0.058
let v9 = if 0 > v8 then 0 else v8
// --- output layer (binary classification) ---
let v10 = 0.739*v7 - 0.415*v9 + 0.022
// sigmoid squashing v10 into the range (0, 1)
let out = 1 / (1 + exp(-v10))
The relu/if-then-else is in fact centrally important as it enables computations with complex control flow (or more exactly, conditional signal flow or gating) schemes (particularly as you add more layers).
My gut tells me that a better approach to a calibrated 0-shot classifier than shoehorning DiffusionGemma would be starting from another gemma, T5GemmaV2. Take its encoder and do continual training on an RTD objective and a large relational synthetic data mix. Then finetuning (multi-annotator data will help calibration) on as many proper NLI datasets as possible. That still lacks the DebertaV3 disentangled attention's inductive bias, however.
Jev also has its calibrated predictions component which is important. Temperature scaling is probably the easiest first pass. But there's lots of sensible options to improve on that.
ModernBERT might be the easier, more stable starting point than T5Gemma though.
reply