Dontopedia
source context

Part 125

ctx:discord/blah/watt-activation/part-125
kind: custommode: permissive149 statements46 distinct subjects

No external document is attached to this context. (Many contexts are pure organisational labels.)

Facts in this context

Grouped by subject. Each subject links to its full article.

Xenonfun18 factsex:xenonfun

asksHow Does That Change Things
attachedScreenshot 2026 03 08 11 50 26 Pm
attachedScreenshot 2026 03 08 11 54 04 Pm
expressesFrustrationNo Checkpoints
expressesPreferenceForTmux With Claude
isLettingRunTo2k Seq Verify Inference
likesTmux Claude Setup
likesConvenienceOfTmux Switching
neverGotCheckpoints
observesTooMuchVocab
postedAt2026-03-09 03:33
postedAt2026-03-09 03:54
postedAt2026-03-09 03:57
postedAt2026-03-09 03:55
postedAt2026-03-09 03:50
sharesObservationVocab Issue
sharesStatusRunning to 2k
statesDesireForSeq 2048

Speaker14 factsex:speaker

assertsMinimalPracticalChangeSeq Increase to 2048
assertsWithCertaintyMinimal Changes
evaluatesAsSafeBs 4 Seq 2048
instructsToUpdateSeq 2048 Bs 4
needsToReCheckBs
offersAlternativeGo With Bs4
offersToMakeChangeAndPushScript
offersToPushChangesScript
offersToRunBs Sweep 4 8 Seq2048
presupposesPriorTestingBs 8 Seq 1024
proposesActionRun Sweep
providesFullPictureMemory Steps Context
seeksConfirmationConfirm or Go With Bs4
summarizesRecommendationUpdate to Seq2048 Bs4

Bs 4 Seq 20488 factsex:bs-4-seq-2048

equalsTokensPerBatch8192
hasIdenticalStepsAndEpochTimeAsBs 8 Seq 1024
hasSameMemoryFootprintAsBs 8 Seq 1024
hasSameStepCountAsBs 8 Seq 1024
hasSameTokSAsBs 4 Seq 1024
limitsBsMaxTo4
losesThroughputAdvantageOfBs 8
maintainsSameTokensAsBs 8 Seq 1024

Training Command8 factsex:training-command

logsEvery100
savesEvery2000
usesBs4
usesCheckpointDirCheckpoints Epoch Cl100k
usesLr1e-04
usesSeq2048
usesWarmup500
validatesEvery2000

Step 1007 factsex:step-100

eta175min
hasLoss11.9294
hasLr2.00e-05
hasPpl151666.9
progress0.6%
stepTime635ms
tokPerS12894

Step 2007 factsex:step-200

eta172min
hasLoss11.4455
hasLr4.00e-05
hasPpl93477.9
progress1.2%
stepTime627ms
tokPerS13067

Step 3007 factsex:step-300

eta172min
hasLoss9.6110
hasLr6.00e-05
hasPpl14927.9
progress1.8%
stepTime630ms
tokPerS13001

Step 4007 factsex:step-400

eta171min
hasLoss8.7389
hasLr8.00e-05
hasPpl6241.2
progress2.4%
stepTime630ms
tokPerS13007

Step 5007 factsex:step-500

eta169min
hasLoss8.2263
hasLr1.00e-04
hasPpl3737.9
progress3.0%
stepTime628ms
tokPerS13038

Lisamegawatts6 factsex:lisamegawatts

advocatesWinningAtLonger Context Even Harder
performedTestsOnMemory Standard
postedAt2026-03-09 03:59
postedAt2026-03-09 04:00
predictsSuperiorityInLonger Context
reportsPersonalTestsMemory Standard Gpt Vs Spectral

Bs 16 Seq 10244 factsex:bs-16-seq-1024

causedMemoryHitOf85GB
hitMemoryUsage85GB
knownToHit85GB
presupposes85gbLimit[object Object]

Model4 factsex:model

learnsLonger Range Dependencies
presupposesHasGatedCumsum[object Object]
presupposesHasPosEmbTo2048[object Object]
trainsOn2k Token Windows

Seq 20484 factsex:seq-2048

changesContext Length
enablesLearningOfLonger Range Dependencies
isFreeUpgradeSame Time Memory Better Context
providesBetterContext[object Object]

Spectral4 factsex:spectral

hasEssentialStabilityMemory Standard
outperformsGpt
presupposesCompetitorToGpt[object Object]
wasStableMemory Standard

Training4 factsex:training

hasSameTokensPerStepBs 8 Seq 1024
presupposesTotalTokensFixed[object Object]
requiresEpochTime~181 min
requiresSteps~16,670

Bs 8 Seq 10243 factsex:bs-8-seq-1024

equalsTokensPerBatch8192
usedMemory~40 GB
wasTested~40 GB

Gated Cumsum3 factsex:gated-cumsum

computesInLinearTimeSequence Length T
handlesLongerSequencesCheaply
hasTimeComplexityO(T)

Bs 8 Seq 20482 factsex:bs-8-seq-2048

causesDoublingOfActivation Memory
hasSameLogitTensorSizeAsBs 16 Seq 1024

Epoch Time2 factsex:epoch-time

approximately181 min
equals181min[object Object]

Spectral Plus Lohe V22 factsex:spectral-plus-lohe-v2

benefitsFromLongerSeqMeaningful
isMeaningfulForLonger Seq

Tmux2 factsex:tmux

allowsSwitchingToOther Console Ctrl B N
presupposesSharedSessionWithClaude[object Object]

Training Progress2 factsex:training-progress

showsDecreasingLoss[object Object]
showsDecreasingPpl[object Object]

Bs 41 factex:bs-4

isSafeDefaultSeq 2048

Bs Increase1 factex:bs-increase

tradesOffWithSeq Increase

Chat1 factex:chat

presupposesModelTrainingContext[object Object]

Chat Log1 factex:chat-log

referencesAttachmentsScreenshots

Checkpoints1 factex:checkpoints

presupposesExpectedButAbsent[object Object]

Cl100k1 factex:cl100k

isCheckpointSuffix[object Object]

Computation1 factex:computation

crunchedFor31s

Eta1 factex:eta

decreasesOverTime[object Object]

Gpt1 factex:gpt

cavedMemory Standard

Inference1 factex:inference

presupposesNeedsVerification[object Object]

Logit Tensor Size1 factex:logit-tensor-size

scalesWithBs Times Seq

Longer Context1 factex:longer-context

hasTeleologicalValueLearning Dependencies

Pos Emb1 factex:pos-emb

supportsUpTo2048

Seq 10241 factex:seq-1024

allowedBsHeadroomTo8

Seq 2048 Bs 41 factex:seq-2048-bs-4

isOptimalTradeoff[object Object]

Sequence Doubling1 factex:sequence-doubling

doublesActivation Memory

Spectral Lohe V21 factex:spectral-lohe-v2

isModelVariant[object Object]

Step Count1 factex:step-count

approximately16670

Tok Per S1 factex:tok-per-s

isStableAround13000

Tok S1 factex:tok-s

roughlySameAsBs 4 Seq 1024

Total Steps1 factex:total-steps

equals16670

Total Tokens in Flight1 factex:total-tokens-in-flight

equalsBs Times Seq

Training Log1 factex:training-log

isFromRealRun[object Object]

Training Run1 factex:training-run

totalSteps16670