# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/49765348

# I built non\-autoregressive decision models with RL a year ago

1250 points · 298 comments

[Full discussion](<https://news.ycombinator.com/item?id=49765348>)

[Read original](<https://laya.convaiinnovations.com/>)

Category: [Research & Evaluation](<https://hacksnap.live/?category=research-evaluation>)

## Skept-o-meter & Hotness

Skept\-o\-meter: High\. Estimated from 35 comments\.

35 comments for the summary\.

Peak rank: \#1

Time in Top 10: 41\.5 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\.

246 recorded rank observations from 2026\-09\-19T18:34:59\.226976\+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 \#12

2026\-09\-30T19:02:02\.291882\+00:00: rank \#12

2026\-09\-30T20:02:49\.392147\+00:00: rank \#13

2026\-09\-30T21:03:42\.990832\+00:00: rank \#11

2026\-09\-30T22:02:11\.294387\+00:00: rank \#12

2026\-09\-30T23:01:52\.722953\+00:00: rank \#12

2026\-10\-01T08:02:24\.797064\+00:00: rank \#12

2026\-10\-01T09:01:35\.899477\+00:00: rank \#15

2026\-10\-01T10:01:12\.485692\+00:00: rank \#14

2026\-10\-01T11:00:47\.508128\+00:00: rank \#14

2026\-10\-01T12:00:42\.477917\+00:00: rank \#12

2026\-10\-01T13:00:51\.124852\+00:00: rank \#12

2026\-10\-01T14:00:59\.01153\+00:00: rank \#12

2026\-10\-01T15:00:51\.048047\+00:00: rank \#13

2026\-10\-01T16:00:53\.034654\+00:00: rank \#14

2026\-10\-01T16:36:37\.180071\+00:00: rank \#15

2026\-10\-01T17:00:23\.382263\+00:00: rank \#15

2026\-10\-01T18:01:44\.046011\+00:00: rank \#16

2026\-10\-01T19:01:47\.420738\+00:00: rank \#17

2026\-10\-01T20:01:39\.585083\+00:00: rank \#16

2026\-10\-01T21:01:45\.872085\+00:00: rank \#17

2026\-10\-01T22:01:37\.214899\+00:00: rank \#16

2026\-10\-01T23:04:11\.395288\+00:00: rank \#19

2026\-10\-02T08:02:26\.819307\+00:00: rank \#20

2026\-10\-02T09:02:51\.51649\+00:00: rank \#21

2026\-10\-02T10:02:12\.346099\+00:00: rank \#21

2026\-10\-02T11:02:22\.794376\+00:00: rank \#21

2026\-10\-02T12:01:14\.681578\+00:00: rank \#21

2026\-10\-02T13:02:29\.511591\+00:00: rank \#21

2026\-10\-02T14:01:40\.936738\+00:00: rank \#20

2026\-10\-02T15:02:30\.225476\+00:00: rank \#20

2026\-10\-02T16:02:47\.060695\+00:00: rank \#19

2026\-10\-02T17:02:31\.351493\+00:00: rank \#21

2026\-10\-02T18:01:22\.271899\+00:00: rank \#21

2026\-10\-02T19:01:23\.681417\+00:00: rank \#21

2026\-10\-02T20:01:33\.801055\+00:00: rank \#23

2026\-10\-02T21:02:55\.453316\+00:00: rank \#26

2026\-10\-02T22:01:50\.233139\+00:00: rank \#27

2026\-10\-02T23:01:20\.268516\+00:00: rank \#25

2026\-10\-03T08:01:30\.186188\+00:00: rank \#26

2026\-10\-03T09:01:57\.91467\+00:00: rank \#23

2026\-10\-03T10:00:55\.732001\+00:00: rank \#22

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

2026\-10\-03T12:01:08\.789384\+00:00: rank \#23

2026\-10\-03T13:00:35\.273559\+00:00: rank \#23

2026\-10\-03T14:00:48\.425606\+00:00: rank \#23

2026\-10\-03T15:01:04\.472963\+00:00: rank \#23

2026\-10\-03T16:01:56\.56175\+00:00: rank \#24

2026\-10\-03T17:01:36\.651036\+00:00: rank \#21

2026\-10\-03T18:01:58\.24207\+00:00: rank \#21

2026\-10\-03T19:01:19\.196589\+00:00: rank \#20

2026\-10\-03T20:00:57\.313759\+00:00: rank \#18

2026\-10\-03T21:01:04\.44439\+00:00: rank \#16

2026\-10\-03T21:20:01\.585706\+00:00: rank \#15

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

2026\-10\-03T23:00:55\.584832\+00:00: rank \#15

2026\-10\-04T08:01:00\.327459\+00:00: rank \#14

2026\-10\-04T09:00:59\.321023\+00:00: rank \#15

2026\-10\-04T10:00:42\.975336\+00:00: rank \#15

2026\-10\-04T11:01:03\.76577\+00:00: rank \#15

2026\-10\-04T12:01:45\.307985\+00:00: rank \#15

2026\-10\-04T13:00:49\.176267\+00:00: rank \#15

2026\-10\-04T14:00:56\.952543\+00:00: rank \#15

2026\-10\-04T15:00:54\.739637\+00:00: rank \#15

2026\-10\-04T16:00:51\.403054\+00:00: rank \#15

2026\-10\-04T17:00:43\.166713\+00:00: rank \#15

2026\-10\-04T18:00:46\.315535\+00:00: rank \#14

2026\-10\-04T19:00:54\.768841\+00:00: rank \#14

2026\-10\-04T20:00:51\.070442\+00:00: rank \#15

2026\-10\-04T21:00:28\.550532\+00:00: rank \#14

2026\-10\-04T22:01:36\.217651\+00:00: rank \#16

2026\-10\-04T23:01:09\.400767\+00:00: rank \#16

2026\-10\-05T08:01:51\.387003\+00:00: rank \#15

2026\-10\-05T09:01:40\.528475\+00:00: rank \#14

2026\-10\-05T10:00:56\.290772\+00:00: rank \#14

2026\-10\-05T11:01:05\.919518\+00:00: rank \#14

2026\-10\-05T12:00:56\.590881\+00:00: rank \#13

2026\-10\-05T13:01:05\.197156\+00:00: rank \#15

2026\-10\-05T14:01:04\.085718\+00:00: rank \#15

2026\-10\-05T15:00:50\.814571\+00:00: rank \#14

2026\-10\-05T16:02:09\.54139\+00:00: rank \#13

2026\-10\-05T17:00:35\.336076\+00:00: rank \#13

2026\-10\-05T18:00:20\.445491\+00:00: rank \#13

2026\-10\-05T19:02:43\.604975\+00:00: rank \#14

2026\-10\-05T20:01:09\.777459\+00:00: rank \#14

2026\-10\-05T21:01:47\.650671\+00:00: rank \#15

2026\-10\-05T22:02:59\.93099\+00:00: rank \#14

2026\-10\-06T08:02:11\.210914\+00:00: rank \#14

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

2026\-10\-06T09:01:21\.688765\+00:00: rank \#16

2026\-10\-06T10:00:58\.047773\+00:00: rank \#16

2026\-10\-06T11:00:41\.689929\+00:00: rank \#17

2026\-10\-06T12:00:41\.7705\+00:00: rank \#18

2026\-10\-06T13:01:24\.509519\+00:00: rank \#18

2026\-10\-06T14:00:26\.950817\+00:00: rank \#18

2026\-10\-06T15:00:49\.066414\+00:00: rank \#19

2026\-10\-06T16:00:58\.166747\+00:00: rank \#21

2026\-10\-06T17:01:17\.009511\+00:00: rank \#21

2026\-10\-06T18:01:43\.791429\+00:00: rank \#22

2026\-10\-06T19:00:46\.048609\+00:00: rank \#21

2026\-10\-06T20:01:14\.687294\+00:00: rank \#21

2026\-10\-06T21:02:13\.414382\+00:00: rank \#20

2026\-10\-06T22:00:40\.926271\+00:00: rank \#19

2026\-10\-06T23:03:26\.867445\+00:00: rank \#20

2026\-10\-07T08:01:41\.339804\+00:00: rank \#21

2026\-10\-07T09:02:38\.026986\+00:00: rank \#22

2026\-10\-07T10:01:35\.069531\+00:00: rank \#22

2026\-10\-07T11:02:07\.770084\+00:00: rank \#22

2026\-10\-07T12:01:14\.996697\+00:00: rank \#21

2026\-10\-07T13:00:47\.45912\+00:00: rank \#20

2026\-10\-07T14:01:18\.307739\+00:00: rank \#20

2026\-10\-07T15:01:57\.541362\+00:00: rank \#22

2026\-10\-07T16:01:28\.474022\+00:00: rank \#21

2026\-10\-07T17:01:15\.036908\+00:00: rank \#22

2026\-10\-07T18:00:53\.300079\+00:00: rank \#21

2026\-10\-07T19:01:49\.456657\+00:00: rank \#22

2026\-10\-07T20:01:59\.032487\+00:00: rank \#22

2026\-10\-07T21:05:36\.655392\+00:00: rank \#24

2026\-10\-07T22:02:27\.105241\+00:00: rank \#25

2026\-10\-07T23:01:17\.452894\+00:00: rank \#24

2026\-10\-08T08:03:55\.083808\+00:00: rank \#23

2026\-10\-08T09:03:58\.68898\+00:00: rank \#21

2026\-10\-08T10:04:05\.364235\+00:00: rank \#23

2026\-10\-08T11:02:04\.410307\+00:00: rank \#22

2026\-10\-08T12:02:36\.39618\+00:00: rank \#22

2026\-10\-08T13:03:09\.206401\+00:00: rank \#22

2026\-10\-08T14:03:30\.113384\+00:00: rank \#23

2026\-10\-08T15:02:54\.444758\+00:00: rank \#21

2026\-10\-08T16:03:13\.067807\+00:00: rank \#20

2026\-10\-08T17:03:47\.92465\+00:00: rank \#19

2026\-10\-08T18:01:50\.360431\+00:00: rank \#19

2026\-10\-08T19:02:55\.719274\+00:00: rank \#19

2026\-10\-08T20:01:36\.359022\+00:00: rank \#18

2026\-10\-08T21:03:34\.000719\+00:00: rank \#18

2026\-10\-08T22:01:13\.459732\+00:00: rank \#17

2026\-10\-08T23:01:25\.318392\+00:00: rank \#16

2026\-10\-09T08:01:53\.314021\+00:00: rank \#17

2026\-10\-09T09:02:59\.93426\+00:00: rank \#17

2026\-10\-09T10:02:22\.652527\+00:00: rank \#16

2026\-10\-09T11:02:04\.807282\+00:00: rank \#16

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

2026\-10\-09T13:02:30\.35034\+00:00: rank \#18

2026\-10\-09T14:01:34\.215134\+00:00: rank \#17

2026\-10\-09T15:01:27\.564392\+00:00: rank \#16

2026\-10\-09T16:02:06\.561056\+00:00: rank \#17

2026\-10\-09T17:03:06\.673648\+00:00: rank \#19

2026\-10\-09T18:01:02\.851052\+00:00: rank \#19

2026\-10\-09T19:01:11\.250291\+00:00: rank \#18

2026\-10\-09T20:01:32\.91907\+00:00: rank \#18

2026\-10\-09T21:02:08\.376144\+00:00: rank \#19

2026\-10\-09T22:01:19\.253075\+00:00: rank \#21

2026\-10\-09T23:01:31\.990912\+00:00: rank \#21

2026\-10\-10T08:01:00\.593376\+00:00: rank \#20

2026\-10\-10T09:01:45\.527975\+00:00: rank \#22

2026\-10\-10T10:00:53\.799353\+00:00: rank \#22

2026\-10\-10T11:00:35\.65851\+00:00: rank \#21

2026\-10\-10T12:01:11\.169647\+00:00: rank \#21

2026\-10\-10T13:00:54\.008377\+00:00: rank \#19

2026\-10\-10T14:01:43\.208766\+00:00: rank \#21

2026\-10\-10T15:01:24\.825188\+00:00: rank \#21

2026\-10\-10T16:01:10\.213328\+00:00: rank \#20

2026\-10\-10T17:00:48\.84189\+00:00: rank \#18

2026\-10\-10T18:00:57\.55132\+00:00: rank \#18

2026\-10\-10T19:00:52\.379463\+00:00: rank \#17

2026\-10\-10T20:00:21\.993912\+00:00: rank \#17

2026\-10\-10T21:01:10\.657738\+00:00: rank \#17

2026\-10\-10T22:00:50\.433569\+00:00: rank \#16

2026\-10\-10T23:01:00\.95027\+00:00: rank \#17

The thread's core dispute is less about whether non\-autoregressive classifiers can work than about novelty, packaging, and zero\-shot usability: Laya is an open, fine\-tunable BERT\-style decision engine with documented limits, while Jev's advantage is presented as a branded, no\-fine\-tune product—though commenters question both the 'breakthrough' framing and Laya's behavior on simple reasoning\.

## The brief

Nandakishor Mukkunnoth recounts publishing a March 2025 arXiv paper and open model for non\-autoregressive, RL\-guided decision models, then says TypeSafe AI's September 2026 Jev launch presented a similar concept as a breakthrough without technical papers, open weights, or open datasets\. He introduces Laya, an Apache\-2\.0 family of bidirectional encoder checkpoints that answer typed choice, score, and noul questions in one forward pass, with a Unicode\-script router for 100\+ languages, claimed sub\-35ms latency, and benchmarks against Jev\. The post also lists limitations: choice accuracy degrades above roughly 20 options, zero\-shot performance is weak without fine\-tuning, and temperature calibration is needed\.

- Laya is presented as an open\-source Apache 2\.0 family of non\-autoregressive decision models with three checkpoints: ModernBERT\-large for English, mmBERT\-base for multilingual use, and a typed\-decisions checkpoint, all answering choice, score, and noul primitives without generating text\.
- The built\-in router inspects Unicode scripts across 22 alphabets and Latin stopword distributions, claims sub\-millisecond routing overhead, and uses preloading to avoid 7–10 second cold\-swap penalties when traffic alternates between languages\.
- The article benchmarks Laya against TypeSafe Jev, claiming typed\-decisions 0\.766 vs 0\.727, AG News 0\.950 vs 0\.910, DAIR Emotion 0\.595 vs 0\.480, ECE 0\.081 vs 0\.246, and P50 latency 32\.8 ms vs 236–276 ms; it notes Jev numbers come from third\-party studies and TypeSafe AI\.
- The author lists limitations: choice questions degrade with more than 20 options, Banking77 scored 0\.425 vs Jev's 0\.870, zero\-shot base models score around 0\.35 on typed\-decisions, fine\-tuning is needed for the 0\.766 result, and temperature calibration cuts ECE from 0\.466 to 0\.081\.
- The post frames Jev as launching the same non\-autoregressive decision concept without technical papers, open weights, or open training datasets, while positioning Laya as the open alternative with 512–1024 token context checkpoints\.

## Discussion themes

### Novelty and prior art are disputed

Commenters question whether Jev or Laya represents a genuine breakthrough\. One says Jev is essentially BERT with more data and predicts labs will replicate it; another says BERT\-style classifiers existed years ago and Jev's launch language overstated novelty\. The OP claims Laya was built on the same research direction a year earlier, but a commenter notes TypeSafe has not published Jev's architecture, so that claim cannot be verified at face value, and another points to Laya's recent GitHub/Hugging Face commits\.

Sources: [Comment 49765997](<https://news.ycombinator.com/item?id=49765997>) · [Comment 49767192](<https://news.ycombinator.com/item?id=49767192>) · [Comment 49765349](<https://news.ycombinator.com/item?id=49765349>) · [Comment 49765889](<https://news.ycombinator.com/item?id=49765889>) · [Comment 49766903](<https://news.ycombinator.com/item?id=49766903>)

### Marketing and productization, not just research, drove attention

Several comments argue Jev's clear branding and launch language explain much of its reception\. One commenter says marketing and branding are as important as the product, and that the OP's single Reddit post with an unclear title failed to communicate the same implications\. Others criticize stealth development and hype, while one says the author's decision to publish papers and weights made the work feel academic rather than productized\.

Sources: [Comment 49769116](<https://news.ycombinator.com/item?id=49769116>) · [Comment 49769332](<https://news.ycombinator.com/item?id=49769332>) · [Comment 49766844](<https://news.ycombinator.com/item?id=49766844>) · [Comment 49766861](<https://news.ycombinator.com/item?id=49766861>) · [Comment 49767192](<https://news.ycombinator.com/item?id=49767192>)

### Technical behavior and limits raise doubts

A commenter flags that Laya's checkpoints have 512–1024 token context versus Jev's reportedly much larger context, calling it a major limitation not surfaced prominently\. Another reports trying Laya on simple die and coin questions and getting nonsensical probabilities, such as 9% for an odd roll with 91% confidence\. A separate commenter tried restating state in Jev and still got confusing numeric results\. Others argue 'can't hallucinate' only means constrained output, not accurate probability, since the model can still be wrong between A and B\.

Sources: [Comment 49766001](<https://news.ycombinator.com/item?id=49766001>) · [Comment 49766988](<https://news.ycombinator.com/item?id=49766988>) · [Comment 49767897](<https://news.ycombinator.com/item?id=49767897>) · [Comment 49768081](<https://news.ycombinator.com/item?id=49768081>) · [Comment 49768152](<https://news.ycombinator.com/item?id=49768152>)

### Zero\-shot usability versus fine\-tuning tradeoff

One commenter says Jev's key advantage is not requiring fine\-tuning: solving an arbitrary classification or routing problem in minutes instead of spending a week creating data, fine\-tuning, and setting up infrastructure\. Another notes Laya's own article says to treat it as a foundation model to specialize, not a zero\-shot oracle, and that they lack time and data to fine\-tune a model for each classifier\. A reply says the interesting part of Jev is in\-context, zero\-setup use, while another argues that whether something is over\- or under\-marketed does not change its function\.

Sources: [Comment 49769607](<https://news.ycombinator.com/item?id=49769607>) · [Comment 49769037](<https://news.ycombinator.com/item?id=49769037>) · [Comment 49767230](<https://news.ycombinator.com/item?id=49767230>) · [Comment 49767722](<https://news.ycombinator.com/item?id=49767722>)

### Deployment and input\-modality constraints matter

A commenter says they want an inference provider if the model cannot run on CPU, because managing GPUs in production is nontrivial\. Another questions Jev's 64k total token request budget and says the waitlist makes it hard to sell internally\. A separate thread asks whether Jev or Laya can classify a support ticket containing a screenshot of a login error; a reply notes Jev is not multimodal, so a screenshot alone would not work\.

Sources: [Comment 49765916](<https://news.ycombinator.com/item?id=49765916>) · [Comment 49766166](<https://news.ycombinator.com/item?id=49766166>) · [Comment 49766740](<https://news.ycombinator.com/item?id=49766740>) · [Comment 49766822](<https://news.ycombinator.com/item?id=49766822>)

### Broader LLM\-versus\-classical\-ML debate

Some commenters say using LLMs for every task is lazy, wasteful, and marketing\-driven, and that experienced ML practitioners know when not to use them\. Others counter that LLMs are less fragile and generalize better on messy classical tasks, and that they ripped out homegrown ML models in favor of LLMs\. A middle\-ground commenter says they use LLMs daily but know when to avoid them\.

Sources: [Comment 49766268](<https://news.ycombinator.com/item?id=49766268>) · [Comment 49766353](<https://news.ycombinator.com/item?id=49766353>) · [Comment 49766457](<https://news.ycombinator.com/item?id=49766457>) · [Comment 49766542](<https://news.ycombinator.com/item?id=49766542>)

## Sources & coverage

AI-generated summary · 2026\-09\-20T10:32:28\.951492\+00:00

Based on 35 of 35 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.