HN user

mycall

1,798 karma
Posts13
Comments1,806
View on HN

Is there not a way to use a transformer that is deterministic in rounds as it is just maths? I thought it was temperature which randomized results but perhaps it is also something else.

Of course, the content returned is non-deterministic.

So the practical workaround is to set model_auto_compact_token_limit very near, but still below, the model’s context window. This postpones compaction rather than disabling it.

config.toml

  model_context_window = 27000
  model_auto_compact_token_limit = 17000
  model_auto_compact_token_limit_scope = "total"

CS101 Day 1: Learn about Divide and Conquer in depth. Honesty forms the ability for controlling signal strength, avoiding impedance mismatches or logically, synchronization between the divided slices.

Qwen 3.8 3 days ago

it also happens to be really good for humanity.

AI being good for humanity is still an open question, but for closed vs. open models/weights, yeah it is preferred. I foresee it won't be much longer before everyone will be slicing/distilling/tuning their models once the architecture improves.

GLM 5.2 vs. Opus 1 month ago

The latest Anthropic and OpenAI models excel in both domains.

Is that because OpenAI models are not a single model but a cluster of models which specialize different domains?

CORS rules only prevent the JavaScript web client from reading the response from a remote endpoint

Assuming the web client plays nicely. I commonly bypass CORS for playwright unit/E2E testing.

Stop Using JWTs 1 month ago

h4ckernews also accessed an Azure Function App that provided direct download URLs for internal FIFA files, including transfer reports and board level data, due to a lack of RBAC access checks.

The biggest problem I see is that web browsers don't natively support .ans inside <canvas> without Ansilove.js or Xterm.js hacks.

Having a native AV format that comes from ANSI, pre-rendered via FFmpeg, is the missing link for <video> support.

Maybe the training approaches taken to date are wrong for decentralized systems. Setup a virtual subnet you can trust and do training on that. Create a AI model island in a trusted/federated model system -- definitely slower than the typical 'one big model' approach, but scalable to world size modeling.

Also, it wouldn't be able to use a transformer architecture. For inspiration, take a look at Google Maps and how it a much more efficient A* divide/conquer hill-climbing architecture. Think minimized matrix math.