# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/49783133

# Show HN: Mini\-AGI – Dynamic continual learning model trained on 8GB VRAM

202 points · 41 comments

[Full discussion](<https://news.ycombinator.com/item?id=49783133>)

[Read original](<https://github.com/volotat/mini-AGI/>)

Category: [Models & Products](<https://hacksnap.live/?category=models-products>)

## Skept-o-meter & Hotness

Skept\-o\-meter: High\. Estimated from 4 comments\.

4 comments for the summary\.

Peak rank: \#8

Time in Top 10: 24\.0 hours

Hacksnap ranks recent stories first, then orders each group by points\. Peak rank uses all retained observations\. Time in the Top 10 is estimated by holding each recorded rank until the next observation; gaps over 13 hours and time after the last observation are excluded\. Movement between observations is unknown\.

236 recorded rank observations from 2026\-09\-21T16:01:23\.189168\+00:00 to 2026\-10\-10T23:01:00\.95027\+00:00\.

Hotness — latest 168 recorded Hacksnap ranks:

2026\-09\-30T18:03:34\.83179\+00:00: rank \#93

2026\-09\-30T19:02:02\.291882\+00:00: rank \#93

2026\-09\-30T20:02:49\.392147\+00:00: rank \#94

2026\-09\-30T21:03:42\.990832\+00:00: rank \#92

2026\-09\-30T22:02:11\.294387\+00:00: rank \#93

2026\-09\-30T23:01:52\.722953\+00:00: rank \#93

2026\-10\-01T08:02:24\.797064\+00:00: rank \#93

2026\-10\-01T09:01:35\.899477\+00:00: rank \#96

2026\-10\-01T10:01:12\.485692\+00:00: rank \#96

2026\-10\-01T11:00:47\.508128\+00:00: rank \#96

2026\-10\-01T12:00:42\.477917\+00:00: rank \#95

2026\-10\-01T13:00:51\.124852\+00:00: rank \#95

2026\-10\-01T14:00:59\.01153\+00:00: rank \#95

2026\-10\-01T15:00:51\.048047\+00:00: rank \#96

2026\-10\-01T16:00:53\.034654\+00:00: rank \#97

2026\-10\-01T16:36:37\.180071\+00:00: rank \#98

2026\-10\-01T17:00:23\.382263\+00:00: rank \#98

2026\-10\-01T18:01:44\.046011\+00:00: rank \#99

2026\-10\-01T19:01:47\.420738\+00:00: rank \#100

2026\-10\-01T20:01:39\.585083\+00:00: rank \#99

2026\-10\-01T21:01:45\.872085\+00:00: rank \#100

2026\-10\-01T22:01:37\.214899\+00:00: rank \#99

2026\-10\-01T23:04:11\.395288\+00:00: rank \#102

2026\-10\-02T08:02:26\.819307\+00:00: rank \#103

2026\-10\-02T09:02:51\.51649\+00:00: rank \#105

2026\-10\-02T10:02:12\.346099\+00:00: rank \#106

2026\-10\-02T11:02:22\.794376\+00:00: rank \#106

2026\-10\-02T12:01:14\.681578\+00:00: rank \#106

2026\-10\-02T13:02:29\.511591\+00:00: rank \#106

2026\-10\-02T14:01:40\.936738\+00:00: rank \#105

2026\-10\-02T15:02:30\.225476\+00:00: rank \#106

2026\-10\-02T16:02:47\.060695\+00:00: rank \#105

2026\-10\-02T17:02:31\.351493\+00:00: rank \#107

2026\-10\-02T18:01:22\.271899\+00:00: rank \#108

2026\-10\-02T19:01:23\.681417\+00:00: rank \#109

2026\-10\-02T20:01:33\.801055\+00:00: rank \#111

2026\-10\-02T21:02:55\.453316\+00:00: rank \#114

2026\-10\-02T22:01:50\.233139\+00:00: rank \#115

2026\-10\-02T23:01:20\.268516\+00:00: rank \#113

2026\-10\-03T08:01:30\.186188\+00:00: rank \#114

2026\-10\-03T09:01:57\.91467\+00:00: rank \#114

2026\-10\-03T10:00:55\.732001\+00:00: rank \#113

2026\-10\-03T11:01:22\.538596\+00:00: rank \#113

2026\-10\-03T12:01:08\.789384\+00:00: rank \#114

2026\-10\-03T13:00:35\.273559\+00:00: rank \#114

2026\-10\-03T14:00:48\.425606\+00:00: rank \#114

2026\-10\-03T15:01:04\.472963\+00:00: rank \#114

2026\-10\-03T16:01:56\.56175\+00:00: rank \#115

2026\-10\-03T17:01:36\.651036\+00:00: rank \#112

2026\-10\-03T18:01:58\.24207\+00:00: rank \#112

2026\-10\-03T19:01:19\.196589\+00:00: rank \#112

2026\-10\-03T20:00:57\.313759\+00:00: rank \#111

2026\-10\-03T21:01:04\.44439\+00:00: rank \#110

2026\-10\-03T21:20:01\.585706\+00:00: rank \#110

2026\-10\-03T22:00:15\.503126\+00:00: rank \#110

2026\-10\-03T23:00:55\.584832\+00:00: rank \#110

2026\-10\-04T08:01:00\.327459\+00:00: rank \#110

2026\-10\-04T09:00:59\.321023\+00:00: rank \#111

2026\-10\-04T10:00:42\.975336\+00:00: rank \#111

2026\-10\-04T11:01:03\.76577\+00:00: rank \#111

2026\-10\-04T12:01:45\.307985\+00:00: rank \#111

2026\-10\-04T13:00:49\.176267\+00:00: rank \#111

2026\-10\-04T14:00:56\.952543\+00:00: rank \#111

2026\-10\-04T15:00:54\.739637\+00:00: rank \#112

2026\-10\-04T16:00:51\.403054\+00:00: rank \#113

2026\-10\-04T17:00:43\.166713\+00:00: rank \#113

2026\-10\-04T18:00:46\.315535\+00:00: rank \#112

2026\-10\-04T19:00:54\.768841\+00:00: rank \#112

2026\-10\-04T20:00:51\.070442\+00:00: rank \#113

2026\-10\-04T21:00:28\.550532\+00:00: rank \#112

2026\-10\-04T22:01:36\.217651\+00:00: rank \#114

2026\-10\-04T23:01:09\.400767\+00:00: rank \#114

2026\-10\-05T08:01:51\.387003\+00:00: rank \#114

2026\-10\-05T09:01:40\.528475\+00:00: rank \#114

2026\-10\-05T10:00:56\.290772\+00:00: rank \#114

2026\-10\-05T11:01:05\.919518\+00:00: rank \#114

2026\-10\-05T12:00:56\.590881\+00:00: rank \#114

2026\-10\-05T13:01:05\.197156\+00:00: rank \#116

2026\-10\-05T14:01:04\.085718\+00:00: rank \#116

2026\-10\-05T15:00:50\.814571\+00:00: rank \#116

2026\-10\-05T16:02:09\.54139\+00:00: rank \#115

2026\-10\-05T17:00:35\.336076\+00:00: rank \#115

2026\-10\-05T18:00:20\.445491\+00:00: rank \#115

2026\-10\-05T19:02:43\.604975\+00:00: rank \#116

2026\-10\-05T20:01:09\.777459\+00:00: rank \#116

2026\-10\-05T21:01:47\.650671\+00:00: rank \#117

2026\-10\-05T22:02:59\.93099\+00:00: rank \#118

2026\-10\-06T08:02:11\.210914\+00:00: rank \#118

2026\-10\-06T08:02:14\.354697\+00:00: rank \#118

2026\-10\-06T09:01:21\.688765\+00:00: rank \#120

2026\-10\-06T10:00:58\.047773\+00:00: rank \#120

2026\-10\-06T11:00:41\.689929\+00:00: rank \#121

2026\-10\-06T12:00:41\.7705\+00:00: rank \#122

2026\-10\-06T13:01:24\.509519\+00:00: rank \#122

2026\-10\-06T14:00:26\.950817\+00:00: rank \#122

2026\-10\-06T15:00:49\.066414\+00:00: rank \#123

2026\-10\-06T16:00:58\.166747\+00:00: rank \#125

2026\-10\-06T17:01:17\.009511\+00:00: rank \#125

2026\-10\-06T18:01:43\.791429\+00:00: rank \#126

2026\-10\-06T19:00:46\.048609\+00:00: rank \#126

2026\-10\-06T20:01:14\.687294\+00:00: rank \#127

2026\-10\-06T21:02:13\.414382\+00:00: rank \#127

2026\-10\-06T22:00:40\.926271\+00:00: rank \#126

2026\-10\-06T23:03:26\.867445\+00:00: rank \#128

2026\-10\-07T08:01:41\.339804\+00:00: rank \#129

2026\-10\-07T09:02:38\.026986\+00:00: rank \#131

2026\-10\-07T10:01:35\.069531\+00:00: rank \#131

2026\-10\-07T11:02:07\.770084\+00:00: rank \#131

2026\-10\-07T12:01:14\.996697\+00:00: rank \#130

2026\-10\-07T13:00:47\.45912\+00:00: rank \#129

2026\-10\-07T14:01:18\.307739\+00:00: rank \#129

2026\-10\-07T15:01:57\.541362\+00:00: rank \#132

2026\-10\-07T16:01:28\.474022\+00:00: rank \#132

2026\-10\-07T17:01:15\.036908\+00:00: rank \#133

2026\-10\-07T18:00:53\.300079\+00:00: rank \#133

2026\-10\-07T19:01:49\.456657\+00:00: rank \#134

2026\-10\-07T20:01:59\.032487\+00:00: rank \#134

2026\-10\-07T21:05:36\.655392\+00:00: rank \#136

2026\-10\-07T22:02:27\.105241\+00:00: rank \#137

2026\-10\-07T23:01:17\.452894\+00:00: rank \#137

2026\-10\-08T08:03:55\.083808\+00:00: rank \#137

2026\-10\-08T09:03:58\.68898\+00:00: rank \#137

2026\-10\-08T10:04:05\.364235\+00:00: rank \#138

2026\-10\-08T11:02:04\.410307\+00:00: rank \#137

2026\-10\-08T12:02:36\.39618\+00:00: rank \#137

2026\-10\-08T13:03:09\.206401\+00:00: rank \#137

2026\-10\-08T14:03:30\.113384\+00:00: rank \#138

2026\-10\-08T15:02:54\.444758\+00:00: rank \#136

2026\-10\-08T16:03:13\.067807\+00:00: rank \#135

2026\-10\-08T17:03:47\.92465\+00:00: rank \#134

2026\-10\-08T18:01:50\.360431\+00:00: rank \#134

2026\-10\-08T19:02:55\.719274\+00:00: rank \#135

2026\-10\-08T20:01:36\.359022\+00:00: rank \#135

2026\-10\-08T21:03:34\.000719\+00:00: rank \#137

2026\-10\-08T22:01:13\.459732\+00:00: rank \#137

2026\-10\-08T23:01:25\.318392\+00:00: rank \#136

2026\-10\-09T08:01:53\.314021\+00:00: rank \#137

2026\-10\-09T09:02:59\.93426\+00:00: rank \#138

2026\-10\-09T10:02:22\.652527\+00:00: rank \#137

2026\-10\-09T11:02:04\.807282\+00:00: rank \#137

2026\-10\-09T12:02:16\.769295\+00:00: rank \#137

2026\-10\-09T13:02:30\.35034\+00:00: rank \#139

2026\-10\-09T14:01:34\.215134\+00:00: rank \#139

2026\-10\-09T15:01:27\.564392\+00:00: rank \#139

2026\-10\-09T16:02:06\.561056\+00:00: rank \#140

2026\-10\-09T17:03:06\.673648\+00:00: rank \#142

2026\-10\-09T18:01:02\.851052\+00:00: rank \#142

2026\-10\-09T19:01:11\.250291\+00:00: rank \#143

2026\-10\-09T20:01:32\.91907\+00:00: rank \#143

2026\-10\-09T21:02:08\.376144\+00:00: rank \#145

2026\-10\-09T22:01:19\.253075\+00:00: rank \#147

2026\-10\-09T23:01:31\.990912\+00:00: rank \#147

2026\-10\-10T08:01:00\.593376\+00:00: rank \#146

2026\-10\-10T09:01:45\.527975\+00:00: rank \#148

2026\-10\-10T10:00:53\.799353\+00:00: rank \#148

2026\-10\-10T11:00:35\.65851\+00:00: rank \#147

2026\-10\-10T12:01:11\.169647\+00:00: rank \#147

2026\-10\-10T13:00:54\.008377\+00:00: rank \#147

2026\-10\-10T14:01:43\.208766\+00:00: rank \#149

2026\-10\-10T15:01:24\.825188\+00:00: rank \#149

2026\-10\-10T16:01:10\.213328\+00:00: rank \#148

2026\-10\-10T17:00:48\.84189\+00:00: rank \#147

2026\-10\-10T18:00:57\.55132\+00:00: rank \#147

2026\-10\-10T19:00:52\.379463\+00:00: rank \#146

2026\-10\-10T20:00:21\.993912\+00:00: rank \#146

2026\-10\-10T21:01:10\.657738\+00:00: rank \#147

2026\-10\-10T22:00:50\.433569\+00:00: rank \#146

2026\-10\-10T23:01:00\.95027\+00:00: rank \#147

The project presents an ambitious, unusually detailed continual\-learning experiment, but the supplied discussion is strongly skeptical: commenters want evidence of generalization, unseen\-data metrics, and coherent outputs, and criticize the AGI framing and missing formal evaluation\.

## The brief

mini\-AGI is a byte\-level continual\-learning language model that trains from scratch on a single 8 GB VRAM GPU by paging expert weights between disk, RAM, and VRAM\. It uses a recurrent MoE architecture with adaptive depth, grows and prunes experts, and claims continual learning without catastrophic forgetting by running the shared trunk at 0\.1x the experts' learning rate\. The README reports early held\-out loss and scaling\-law results but notes the model is toy\-level and weights are not yet published\.

- Architecture: byte\-level vocabulary of 265 tokens, context 4,096, RMSNorm/RoPE/SwiGLU/flash attention, 3 distinct blocks applied up to 26 times per character, a recurrent block applied up to 24 times, and PonderNet\-style adaptive halting\.
- Memory design: expert weights and Adam moments are stored as files on disk and paged into RAM/VRAM as needed, so parameter count is bounded by disk space rather than VRAM; caches and hysteresis limit churn\.
- Growth and pruning: the expert pool grows on speculation and prunes experts that stop being addressed; new experts are built by recombination, and the README argues gate values are anti\-predictive, so 'dead' means unaddressed rather than low\-gate\.
- Continual\-learning claim: single\-stream batch\-1 training with the trunk at 0\.1x the experts' learning rate reportedly reduces forgetting from \+2\.2300 to \+0\.0067 nats in a chess probe, retaining 99\.84% of progress against chance\.
- Benchmarks and scaling: at 318\.1M characters read and 169 experts, held\-out loss across eight subjects is 0\.8336 nats/char (1\.2026 bits/byte); the fitted power law is L ∝ D^\-0\.239 with R² 0\.96\.
- Usability: users can point train\.py read at local files or directories, with \-\-save required to keep changes; the reference machine is an RTX 3070 Laptop GPU with 8 GB, and the run is still reading its first corpus pass\.

## Discussion themes

### Generalization vs\. memorization

A commenter asks whether the architecture can actually generalize or is mostly based on memorization, and suggests basic generalization tasks such as number addition\.

Sources: [Comment 49784604](<https://news.ycombinator.com/item?id=49784604>)

### Unseen\-data metrics requested

Another commenter asks what character prediction rates the model gets on unseen datasets, pointing to a missing quantitative evaluation\.

Sources: [Comment 49785536](<https://news.ycombinator.com/item?id=49785536>)

### Transcript shows no coherent responses

A commenter says that scrolling through the linked transcript shows something that looks like training but no coherent responses at any point, which may explain the absence of benchmarks\.

Sources: [Comment 49786083](<https://news.ycombinator.com/item?id=49786083>)

### AGI framing and missing methodology criticized

A professor publishing on continual learning criticizes the use of 'AGI' without performance analysis, calls it massive marketing overreach or slop, and says the project lacks related work, algorithm description, and ablation study; the 'how\-continual\-learning\-works' text is not an algorithm\.

Sources: [Comment 49786171](<https://news.ycombinator.com/item?id=49786171>)

## Sources & coverage

AI-generated summary · 2026\-09\-21T16:01:06\.944509\+00:00

Based on 4 of 4 usable stored comments, selected by depth and branch activity. This is a sample of the discussion. Article text may also be shortened.

Generated using deepseek\-ai/DeepSeek\-V4\.1\-Flash. Check the linked sources for full context.