the digit 1 in COVID-19
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 2 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron fires on the number 1 in 'COVID-19', the one spot in the name where the model has to work out the rest of the word. Turning it up did not make the model's replies about COVID.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
to eggs.⏎Note, however, that the flu vaccine will not protect against COVID-19, and the COVID-19 vaccine will not protect against flu. Moreover
-
during the pandemic. Organizations and prescribers are under even greater pressure to help ensure patients adhere to their medication regimen. Many hospitals are overwhelmed with treating COVID-19 patients. If a medication nonadherence event requires a patient to receive emergency
-
countries and we expect it to soon be the dominant COVID-19 strain circulating worldwide.⏎“The world is watching in real-time as the COVID-
-
should be well informed about the disease-causing virus and the ways through which it spreads the disease. The only easy and effective way to protect yourself from COVID-19 is by using an alcohol-based massage or washing your hands frequently, and not
-
even more hesitant to leave their homes to go to the pharmacy. But it is particularly important that such higher-risk patients maintain adherence to reduce their vulnerability to COVID-19.⏎- Public mail carriers will typically leave packages in a mailbox or at a
-
related to COVID-19.⏎A major social determinant worsening these health inequities for patients with disabilities
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: I have been a small town in a small town in a small town in a small town in a small town in a small town. The number of people in the small town is 100. The number of people in the small
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
974 .. to describe a global TCP/IP network. So, you could also argue the Internet (<|endoftext|>このトピックには0件の返信が含まれ、1人の参加者がいます。2020年7月6
-
(var1 x)⏎> that changed, and there may be nothing about var11 . var1 $ x that⏎> changed, etc. Yet one is apparently forced to keep track of and⏎> reconstruct all those elements.⏎The compiler
-
vaccinations for workers and patients⏎04 - Measures to prevent central venous catheter-related infections⏎05 - Measures to prevent ventilator-associated pneumonia (nosocomial pneumonia)⏎06 - Measures to prevent surgical site infections⏎07 -
-
the state is one of the nation’s leading producers of biofuels, pumping out over 1.2 billion gallons per year.3⏎However, all is not well on the policy front. Indiana also has a renewable portfolio standard, but
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L1:2670 suppresses
- L1:3043 weakly excites
- L0:2747 weakly suppresses
- L1:1104 weakly excites
- L0:557 weakly excites
- L0:469 weakly suppresses
- L0:2730 weakly excites
- L1:2922 weakly excites
- L1:3046 weakly excites
- L1:146 weakly excites
- L0:2690 weakly suppresses
- L0:1034 weakly excites
- L0:1091 weakly suppresses
- L1:198 weakly suppresses
- L1:2484 weakly suppresses
- L1:2620 weakly suppresses
- L0:2940 weakly excites
- L1:1912 weakly excites
- L0:2418 weakly excites
- L1:549 weakly excites "g" in "e.g." examples
- L0:2811 weakly excites
- L0:338 weakly excites
- L1:2240 weakly suppresses
- L1:1238 weakly excites
Feeds later layers
- L4:1242 suppresses
- L7:767 weakly suppresses
- L13:1166 weakly suppresses
- L7:2132 weakly suppresses
- L16:1072 weakly suppresses
- L20:49 weakly suppresses
- L17:1789 weakly suppresses
- L15:2100 weakly excites
- L14:2012 weakly excites
- L5:286 weakly excites
- L13:958 weakly suppresses
- L26:237 weakly suppresses
- L6:2935 weakly excites
- L15:1804 weakly excites
- L5:2988 weakly suppresses
- L24:340 weakly excites
- L23:643 weakly excites
- L9:1069 weakly excites
- L7:1958 weakly suppresses
- L21:14 weakly suppresses
- L3:1375 weakly excites
- L16:2107 weakly suppresses
- L22:1044 weakly excites
- L7:249 weakly suppresses
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- shit
- osphate
- },{"
- urse
- ัส
- cử
- Duterte
- js
- souvenir
- 放弃了
Suppresses
- 界面
- neys
- omid
- ney
- sthrough
- linked
- competitive
- ouver
- ekte
- oup