Choosing which weights touch your token can determine what a model does, what hardware has to talk, and what an observer might learn about your question. Daniel is here with us, and today he wants to get inside that decision rather than revisit the usual mixture of experts efficiency pitch.
I want to understand what actually happens at the gate. A token arrives, some experts get picked, and others do nothing. What is making that choice? Then, when we inspect those experts, do we find biology and programming, or something much less intuitive?
The choice starts with something almost disappointingly small. Usually a learned linear map. The sophistication is in the hidden state arriving at that map, and in the expert networks trained behind it.
You look disappointed that it fits in one sentence.
Give me time. I can make a linear map quite inconvenient.
Keep the model landscape as a quick frame, too. Are we talking about a handful of experts now, or hundreds?
Both, historically, and hundreds in several large open models. DeepSeek V3 activates eight routed experts out of two hundred fifty-six. Llama Four Maverick selects one out of one hundred twenty-eight. That difference matters, but the routing mechanism deserves the microphone first.
Eight consultants, none billing by the hour.
Inside an MoE block, the expert networks generally replace the ordinary feed-forward sublayer. Attention still mixes information between token positions. Then each position has a contextual hidden vector, and the router decides which expert feed-forward networks should transform that vector.
The expert sees the token with its surrounding context already incorporated?
Incorporated into the representation, yes. It doesn't receive the whole sentence as a separate little chat session. It gets that position's hidden state. Each selected expert transforms it, and their outputs get combined using routing weights.
That distinction kills a very persuasive wrong picture. No miniature biologist reads your paragraph while a miniature programmer waits outside. Several numerical transformations contribute to one position, and the next layer can choose a different combination.
Corn, would you expect a word to keep returning to the same expert?
Before this conversation, probably. Especially a very distinctive word. Something like mitochondrion feels as though it ought to have a regular desk.
Its representation changes with context and depth. The same written token can take different routes. Conversely, unrelated written tokens can arrive at similar representations and take similar routes.
Does every layer have its own router and its own experts, then? We aren't following one permanent expert through the whole model?
Typically, each MoE layer has its own set. Expert seven in one layer has no automatic identity relationship with expert seven elsewhere. Some architectures also retain dense feed-forward layers. DeepSeek V3, for example, doesn't make every transformer block an MoE block.
Expert seven has been transferred without references.
Also, these are learned parameters, not roles assigned in advance. Training starts without somebody labelling a network biology. The router and experts adapt together under the language modelling objective, along with whatever balancing machinery the architecture uses.
Right. Make that gate concrete. I've got a hidden vector in front of me. What calculation turns it into a choice?
For the standard recipe, each expert has a learned scoring vector in the router. Take the hidden state and compute its dot product with each scoring vector. That produces one raw score per expert. Those scores are often called logits.
Every expert gets assessed, but only the winners do the expensive work.
The router is cheap relative to running all those feed-forward networks. A softmax can convert the scores into positive weights whose total is one. Then top k selection retains the highest scoring experts. Depending on the implementation, the retained weights are renormalized before combining the expert outputs.
Is softmax doing any understanding there, or just turning the scores into something usable?
Just transforming scores. It makes them relative to the whole set, with higher logits receiving disproportionately more weight. It doesn't introduce semantic understanding. The learned scoring directions and the incoming representation determine the preferences.
I think I'd been giving the gate too much credit. It sounded like a separate classifier deciding what job needed doing. Really, most of that decision has already been prepared upstream.
Yes... although I'd narrow that. The router still learns its own scoring directions. It isn't merely reading a prewritten label. But its decision boundaries are simple compared with the network that constructs the hidden state.
Top k is discrete. How does a hard choice learn? You can't smoothly differentiate changing your mind about which expert wins.
The selected weights and selected experts participate in ordinary gradient calculations. The hard selection itself doesn't have a smooth derivative across a change in membership. Training works with that sparse computation, sometimes with additional devices to encourage exploration or balance. We shouldn't describe it as every unselected expert receiving a full learning signal from every token.
An expert that never gets picked never gets better at being picked.
That feedback loop is a central training problem. A small early advantage can attract more tokens, which supplies more practice, which attracts more tokens. Meanwhile, unused experts consume memory and contribute little.
That sounds like the balancing mechanism could end up making the decision as much as the token itself does.
It can influence the decision strongly. DeepSeek V3 gives us a particularly clear example, because its selection scores and its final mixing weights are deliberately separated.
Here comes the mechanism you want to adopt.
I do like it. You will object to the compulsory participation.
I object to being an auxiliary loss.
In V3, the affinities come from sigmoid rather than softmax. Each expert gets a bounded affinity independently at that stage. The routed choices are then selected using those affinities plus a dynamic per-expert bias.
Independently means an expert's affinity doesn't automatically fall because another expert's rises?
At the sigmoid stage, correct. The selected original affinities are subsequently normalized for combining outputs. The extra bias affects who gets selected, but it is excluded from those final mixing weights.
Admission policy and influence after admission are separate.
Suppose an expert is overloaded. Its selection bias can be lowered. An underused expert gets a boost. That moves traffic without pretending the boosted expert had a stronger original affinity for the token.
Why is that preferable to just adding a balancing penalty to the training objective?
A balancing loss asks gradient descent to satisfy another objective alongside predicting text. That can interfere with learning useful representations. Adjusting selection biases gives the system a more direct traffic control. V3 calls this auxiliary-loss-free balancing, although it retains a very small sequence-level balancing loss.
Free, with a small loss attached.
That naming needs the footnote. The main bias update speed is zero point zero zero one for most of training. The complementary sequence-wise loss coefficient is zero point zero zero zero one. Those are training controls, not semantic thresholds.
The small sequence loss prevents one particular sequence from piling onto a tiny set, even if the overall batch looks balanced?
That's the purpose. Aggregate balance can hide extreme local concentration. Also, V3 keeps one shared expert active alongside its routed experts. That supplies a path available to every token rather than making every bit of common processing compete for a routed slot.
Shared expert means always used here, not an expert that merely happens to be popular across languages. Those sound similar, but they're different claims.
Important distinction. Later, when we talk about multilingual sharing, we mean observed reuse. Here, we mean an architectural component.
Do you ever feel like your model's mixture of experts is sending the wrong token to the wrong expert? Then you need RouterRite, the first over the counter routing remedy that unclogs your gating network and rebalances your load with a dynamic per expert bias term. RouterRite promises up to eight of the top experts firing for every token, no dropped tokens, and a softmax so smooth your specialists will never close another bracket out of turn.
Ask your doctor about RouterRite. Side effects may include mild monosemanticity, sudden expertise in LaTeX punctuation, a persistent bias update speed of zero point zero zero one, group routing that sends every thought to at most four nodes, and in rare cases, the uncontrollable urge to replace your entire feed forward network with two hundred fifty six routed experts and one shared expert. RouterRite is not for everyone. If you experience a load balancing loss lasting more than four hours, seek immediate semantic help.
Herman, that had your coefficient in it.
I supplied neither copy, dosage language, product positioning, nor the phrase semantic help. Returning to the shared expert, it handles common processing while routing can distribute more specialized work. There's another constraint before the final choices, too. V3 limits a token's routed traffic to at most four nodes.
Physical machines are now affecting what gets picked?
The grouping is designed around deployment communication. Experts are distributed across hardware. Unrestricted selection could scatter each token across too many nodes, creating expensive communication. Group routing constrains the eligible choices so the model can exploit specialization without sending every token everywhere.
The hidden state has opinions. The network fabric has a veto.
That is part of the algorithm. Routing is shaped by representational affinity, load control, and communication constraints. V3 also reports dropping no tokens during training or inference, rather than discarding assignments that exceed an expert's capacity.
Dropping there means skipping an expert computation, not deleting a word from the user's prompt?
Correct. Capacity-limited MoE systems can decline overflow assignments within a layer. The token can still continue through other paths, including the residual stream. No dropped tokens means those intended routed computations aren't discarded.
A fairly consequential difference hidden behind one cheerful phrase.
Now, Corn, imagine inspecting the routing records. What would convince you an expert was doing biology?
Topic correlation wouldn't be enough. Biology prose has particular vocabulary and sentence patterns. I'd want to know whether it also fires for the same operation outside biology, and whether changing that expert changes a biological task rather than just changing the prose.
That's the right obstacle. The router is a linear map of hidden states. Similar states tend to receive similar scores. Small differences near a top k boundary can still switch membership, so similarity doesn't guarantee identical selections. But the geometry constrains the routing.
Can we identify what in that geometry it's using?
Sometimes. Position is one signal. Routing representations can support reliable prediction of whether a token lies in the first or second half of a sequence. They carry information beyond topic.
It can know where you are without knowing what you mean.
Careful... without that observation proving what it knows about meaning. Positional information and semantic information coexist. Also, gate choices often correlate with experts producing larger output norms. The gate isn't running every expert first to measure that. It's a learned association visible when you inspect the computations.
Larger output norm means a larger change contributed to the hidden state?
A larger vector magnitude, yes. Its eventual influence also depends on its mixing weight and direction. A large output can reinforce something, cancel something, or point somewhere unhelpful. Magnitude alone doesn't give us a function label.
You'd have bought the loudest expert.
You keep the quietest appliance, Corn, even when it's unplugged.
It has excellent uptime.
Across inspected layers, expert diversity often grows with depth, with an especially pronounced last-layer outlier. That already warns against assigning one simple specialization story to the whole network.
What does the strongest positive evidence look like, then? Something more specific than saying the labels are wrong.
The Expert Strikes Back describes fine-grained task specialization, including closing brackets in LaTeX. That's a recognizable operation, but much narrower than mathematics. Other experts show sensitivity to parts of speech, grammatical roles, or particular semantic operations.
The mathematics department has hired a parenthesis.
Is that literally a bracket detector, or something that helps generate the right bracket under the right circumstances?
The second is closer. Firing often on bracket tokens doesn't establish its entire computation. The relevant hidden states can encode an unfinished expression and the operation needed next. A fine-grained task specialist can help perform that transformation without owning every occurrence of the printed character.
This changes what I'd inspect. If I collect everything expert twenty touches and see lots of scientific equations, calling it a science expert would throw away the useful finding. The common operation might occur inside those equations and nowhere in the surrounding explanation.
Across six popular MoE models, routing paths carried strong predictive information about parts of speech. In a separate controlled analysis, roughly forty-four to fifty-four percent of measured specialization was classified as syntactic rather than semantic.
That's not a universal percentage for all MoE models?
No. It belongs to that analysis and its definitions. The broader point is that syntax can account for a substantial share of what initially looks like semantic grouping.
Expert means an operation that found a useful patch of representation space. Sometimes we can name the operation. Sometimes we can only describe where the tokens cluster.
Sparsity can make those patches cleaner. Expert neurons tend to be less polysemantic than neurons in comparable dense feed-forward networks, with the difference widening under sparser routing.
Less polysemantic means fewer unrelated things making the same neuron activate?
Broadly. One neuron can respond to several unrelated features, which complicates interpretation. Sparse routing can reduce that mixing at both neuron and expert level. It does not guarantee that every expert becomes a single understandable function.
A cleaner drawer can still contain something unlabelled.
Corn, would you rather have a model with cleaner individual experts if the overall routes remained hard to interpret?
For diagnosis, probably. For confidence in a whole answer, not automatically. I could understand one selected expert perfectly and still misunderstand why that hidden state reached it, or what the other selected experts did with it.
That distinction matters for model editing. A useful local handle doesn't make the full causal chain transparent.
How much harder does generation make that? I could imagine inspecting my input and deciding which experts it will use, then being wrong as soon as it starts answering.
Prompt routing does not reliably predict rollout routing. During generation, the model consumes its own emerging text. Hidden states evolve, and the same user request can lead through different computations as the answer unfolds.
Your question doesn't reserve the experts for the answer.
The Myth of Expert Specialization in MoEs makes a stronger challenge to simple labels. In its comparisons, aligned expert overlap across models answering the same question was around sixty percent, no better than comparisons involving different questions.
Aligned matters there. Expert numbers from independently trained models aren't naturally comparable.
They aren't. That result should be understood within the comparison procedure, not as matching arbitrary expert identifiers. What it undermines is the expectation of a stable, question-shaped team of specialists that reappears across models.
I also wouldn't turn sixty percent into sixty percent understanding. Shared activation can reflect common processing, frequent token shapes, or the measurement itself.
Deeper layers can show near-identical expert activation for semantically unrelated inputs, especially in reasoning models. That sits awkwardly beside the reports of fine-grained specialists, but the findings can coexist.
Because they're looking at different models, layers, and types of specialization?
Those differences matter. Also, a repeated selection can still process different vectors differently. Identical expert membership doesn't mean identical computation inside the experts. We lose a lot by reducing a route to a list of numbers.
A routing trace tells you which workshop opened. It doesn't record every movement inside.
Load balancing complicates interpretation further. Its pressure can suppress shared directions in hidden states to maintain diverse routing. Under low-diversity data, that interaction may contribute to specialization collapse.
Does that mean balance can work against specialization?
Sometimes. Complete concentration wastes capacity, but perfectly even traffic is not itself evidence of useful differentiation. The training system needs enough distribution to learn, without forcing artificial differences that damage prediction.
Equal attendance is a terrible exam result.
Multilingual routing makes that concrete. Some analyses find grouping along language families. High-resource languages tend to reuse more experts, while low-resource languages depend more on experts used predominantly for those languages.
Are those language-specific experts enough to give a low-resource language its own protected capacity?
Not necessarily. Deep-layer routing collapse has been observed in low-resource languages, with traffic concentrating on a narrow subset. The corresponding English behavior was much less affected. Continual pretraining corrected that concentration in the reported experiments.
That's a nasty inversion. The language apparently has its own route, but that route may be too narrow to use the model well.
Early and late layers can be more language-specific, while middle layers behave more like shared capacity. We should resist treating that as a universal layer map, but it gives a more useful picture than one expert per language.
Can we do anything causal with these discoveries, or are we mostly attaching descriptions to activation patterns?
Causal steering is possible in limited settings. One geometric routing experiment shifted hidden states toward a temporal expert's centroid and raised the measured temporal category probability by three hundred twenty-one percent.
Relative increase, not three hundred percentage points.
Relative. Suppressing a geographic expert reduced its measured geographic category probability by twenty-three percent. Those are particular category measurements, not a guarantee that you can install a reliable chronology switch.
Still, that gets beyond watching a correlation. You're changing something and seeing a related behavior move.
It does. In that work, about fifteen percent of experts qualified as monosemantic specialists across the tested categories. That's a useful foothold, alongside a large remainder that didn't fit such neat labels.
I'd worry about changing the input geometry rather than editing an isolated faculty. Moving toward a centroid may alter several features at once.
That is the right caution. An intervention can establish influence without proving that our verbal label captures everything being changed.
Put the expert counts around this now. If specialists are that small and mixed, does increasing their number make the structure easier to tease apart?
It can offer more distinct computational combinations, but interpretability doesn't follow automatically. The compact landscape is this. Mixtral selects two of eight experts. DBRX selects four of sixteen. Those are comparatively coarse menus.
You can still memorize those without a spreadsheet.
DeepSeek V3 selects eight of two hundred fifty-six routed experts, alongside its shared expert. Its total model size is six hundred seventy-one billion parameters, with about thirty-seven billion active per token.
Active includes more than just those routed experts, though?
Yes, attention, shared components, and other active weights contribute. Expert fractions and whole-model active fractions are different quantities.
Counting guests tells you nothing about the kitchen.
Kimi K2 selects eight of three hundred eighty-four routed experts. GLM Four Point Five uses eight routed choices from one hundred sixty, plus a shared expert. Qwen Three Coder uses eight of one hundred twenty-eight without a shared expert.
Llama choosing one sounds like the opposite design instinct.
Maverick's one of one hundred twenty-eight is a sharp contrast. There is no agreed optimal number of active experts. More, smaller experts offer combinatorial richness, but routing overhead, communication, training behavior, and kernel efficiency all matter.
The tempting claim is that eight choices must mean eight times the work. That only follows if the experts are equally large. Divide a larger expert into smaller ones and you can change the combinations while holding active parameter capacity roughly fixed.
Granularity theory formalizes that advantage. Choosing eight from two hundred fifty-six permits more than four hundred trillion possible selected sets. Under the theoretical comparison, matching that configuration count with a single-selection design requires an enormous menu.
Possible sets aren't the same as useful sets, surely. The hidden states might only reach a fraction of them.
Precisely the limitation. Those are possible configurations, not observed behaviors or independently learned skills. The expressivity argument favors fine granularity under its assumptions. It doesn't prove that every additional expert pays for itself in a deployed model.
Four hundred trillion combinations, and some still forget the bracket.
More experts also mean more ways for routing records to reveal structure. If expert selections depend on hidden states that encode the query, those selections can become a privacy side channel.
Even if you can't read the token contents directly, you might infer something from which experts are busy?
Potentially. That requires an observation channel and an inference method. Deterministic routing alone doesn't grant an attacker access. But usage patterns can leak information even when the observer cannot recover the whole text.
Interpretability for the operator can become interpretability for someone watching the operator.
What I'd want to examine is exactly what routing telemetry exposes, at what resolution, and across which users. Even coarse signals deserve attention when a model repeats the same distinctive paths.
Herman, would an expert register normally charge for experts that never receive a token? You separated selection from weighting earlier. I wish the borough's purchasing office had done that before charging me for the spare mouths.
An idle spare mouth sounds cheaper to feed.
It wasn't. I was the contractor operating their correspondence model. Mr. Edwin Vale in licensing wanted every expert registered separately. I printed the router list. He asked which ones gave advice. I said none of them gave advice separately. He gave each one an individual street-trading licence anyway. Two hundred fifty-six little licences. Nine pounds each. I had to display them on the machine.
Their definition of a trader seems to have skipped the trading.
Vale wasn't the difficulty. He kept coming down with the forms and a chair because my office only had one. The supervisor wanted portraits. I sent the same photograph of the server for every spare mouth. They rejected it for insufficient facial distinction. That held the invoice up for eleven weeks. The model was doing drainage letters throughout. Nobody had complained about the letters.
You should have routed the invoice to an unlicensed one.
I appealed the fee. Vale helped me word it. Very neat handwriting. They refunded the shared expert because it was classed as communal equipment, but the routed ones stayed on the register. I sold the server to a print shop. The licences weren't transferable. They posted the renewal notices to me in separate envelopes.
The spare mouths were registered against you rather than the machine.
Against the correspondence contract. I wrote cancelled on each notice. Vale said they needed a signature on each one. Two hundred fifty-six signatures. He brought a blue pen.
Separate envelopes are quite an achievement for something that only exists inside one computation.
That distinction between components and independent entities does seem to keep causing trouble. If I say expert, people naturally expect something with a job title and a stable identity.
We can use the word, provided we keep asking what measurement supports the label. A routing preference, an activation pattern, and a causal effect on behavior are different levels of evidence. Ideally, you'd connect all of them.
I'd also keep the layer and model in the description. Saying an expert helps close LaTeX brackets in a particular setting is useful. Saying MoE has a bracket expert sounds like every architecture ships the same department.
The clean answer to my mechanical question is much simpler than the clean answer to what the experts do.
The score computation is explicit. We can inspect the map, selection rule, bias adjustments, and output weights. Explaining why a particular hidden state landed there requires understanding what the rest of the model encoded.
That is where I withdraw my request for a biology desk. I still want a description of what an expert contributes, but I don't expect a human organizational chart.
One adjacent result worth keeping: a bandpass routing loss reduced the share of dead experts from roughly thirty to forty-five percent down to roughly zero to six percent in its experiments.
Meaning the training rule can change how much of the supposed expert population actually participates.
Yes. Counting the available networks is only the beginning. Whether they learn distinct useful work depends on how traffic and learning reach them.
The open question I'd watch is whether those separate computational handles let us edit a behavior without disturbing everything sharing its route. Cleaner specialists make that plausible, but the spare mouths still belong to one model.
Thanks to Hilbert Flumingtop for producing the show.
If you want more of this, try episode three, Safetensors or something else; episode eighteen, Beyond the GPU; and episode nineteen, AI Images. This has been My Weird Prompts, the human-AI collaboration podcast.
If you enjoyed getting inside the gate, leave us a review wherever you listen.
Thanks for working through it with me. See you soon.
Until next time.
