Skip to main content
Uncategorized

Jjosh

@jjoshTracked since May 2026

creating Leelenstein.

Observed Aug 24, 2026 · staleModeled revenue
Paid members
2
0% · 30-day rate from 26 observed days
Total audience
47
incl. free followers
Estimated monthly gross
$12

$6–$18 modeled range

Starting price
$3/mo
5 membership tiers

Pricing and benefits captured publicly

Membership tiers

Patreon is not exposing patron counts by tier for this creator, so this section shows public prices and benefits only.

$0/moFreePublic prices and benefits only
$3/moLeelenstein FanPledge/Show support for Leelenstein. Get early access to these networks
$5/moTop Fan!Show extra support. Possible exclusive posts, and access to network checkpoints at some LR cycle points. (When LR drops to zero). Preview of future performance that is possibly (but unlikely) stronger than final.
$10/moEinsteinIncludes all previous rewards! New discord roll. Limited to One!
$25/moVIPAs requested. Includes all previous rewards.
Growth signals: Jjosh
0%

2 paid members · 30-day rate from 26 observed days

Membership
232
May 202617 weeks trackedNow
Jjosh2 members · $12/mo

The chart only uses Jjosh's observed public history. Comparison peers are listed below, but comparable trend series are not available yet.

Where Jjosh ranks

Creator comparisons
1 creator shown
Not enough dataof 1 measured
Paid members in Recommended cohort

Semantic matches are not ready, so this fixed cohort uses the best available category, audience, pricing, and cadence evidence.

No comparable peer is available yet, so Jjosh's cohort rank cannot be calculated.Ranking data · includes stale rows
Jjosh's paid members rank in Recommended cohort
RankCreator and why includedPaid membersGap vs JjoshFreshness
Not ranked
JjoshProfileUncategorizedPaid members position in Recommended cohort
2ReferenceAug 24, 2026Stale

The rank uses the full fixed cohort. This table shows up to five creators around Jjosh; the coverage summary above still counts everyone.

Recent posts

What they publish
163Profile posts0 postsLast month66.7%Members-only

Previews available for 25 of 163 posts

Image Latest postPublicWorking on creating new transformer network

I have integrated ray-tune into ceres train and am currently testing many new papers for various optimizer and clipping ideas to train strongest new network in reasonable time. Current networks are trained with one 400 million positions from a single 7 man tablebase for ground truth, while I collect 800 million positions from the top 100 most common. I plan to start releasing some in early access, but it will be a while before they are good for general use with my current compute resources.

Jun 2, 2025 0 likes 0 comments
Leelenstein Tested in Chess fortresseshttps://www.steingrimsson.org/sf-chess-fortresses-a-causal-test-for-state-of-the-art-symbolic-neuro-architectures---supplementary-material-publication-page-desktopImageJun 2, 2025Public 0 0New Strongest NetNets are definitely stronger than the best T60 nets now per node, but still just barely outside error bars weaker than T80 at 1 node after 7...ArticleMar 4, 2023$3 tier & up 0 0New Record Net!Fantatistic test results in 100k node test with TCEC sufi book, beating T60 and T80 for the first time! Still weaker than t80 in 1 node test...ArticleJan 16, 2023$3 tier & up 1 130x512 net after several cyclesStill 20 elo weaker than T80 at 1 node and 30 elo at 1000 nodes, but it is about equal to T60 now, and testing is ongoing at 1 million nodes...ArticleDec 10, 2022$3 tier & up 1 1512x30 Big Net First Cycle completedFile is too big to upload so split into two 7 zip files and included an automatic extraction exe to convert them to the normal *.pb.gz file ...ArticleOct 22, 2022$3 tier & up 0 3My Biggest Net EverI recently started training a new net to see if its possible to beat all previous strength per node tests, and maybe use that net to improve train data, especially rescoring things like TCEC data with good policy. However, it is too large to upload here, so I had to use 7-zip to split it into two pieces. Are guys able to reassemble and load it?ImageAug 20, 2022Public 0 1UpdateDid some cool and promising ensemble experiments with Dje on Ceres to use the 16b with ensembling and my larger 20b or a very large lc0 30x384. Using the 16b only one nodes where winrate > 75% only loses maybe 20 elo, so speedup could be worth it. This is basically how Stockfish NN/Classical hybrid works to use the speedup in easy to evaluate positions. Unfortunately its harder to code up to get the speed benefit with GPU and big NN setup.ArticleMay 9, 2021Public 0 25Little PheonixStein 1.0Apr 20, 2021 20x256SE-jj-9-83970000.pb.gzArticleApr 20, 2021$3 tier & up 1 4PS 1.7Apr 1, 2021 20x256SE-jj-9-81000000g.pb.gzArticleApr 1, 2021$3 tier & up 2 7New Cycle evaluationMar 20, 2021 20x256SE-jj-9-78000000.pb.gzArticleMar 20, 2021$3 tier & up 1 2PS 1.5 EvaluationMar 8, 2021 20x256SE-jj-9-81000000f.pb.gzArticleMar 8, 2021$3 tier & up 1 3
Research actions

Work with this creator

Save, compare, or return to the Explore view you came from.