#serverless — Public Fediverse posts on home.social

Sascha Fedorenko 🇨🇭🇺🇦 @[email protected] · 2026-09-29 · 10:00 UTC

https://the-well-architected-cloud.com/blog/do-androids-dream-of-flappy-birds/

TypeSafe released Jev, a "decision model" that's "just like an LLM, but not an LLM". They didn't share how it works.

So I guessed. My hypothesis: it's a regular LLM that stops before generating text and reports how likely each option is to continue the prompt.

To test it, I built Jeb: an unmodified LLM, llama.cpp and the raw next-token probabilities. It works. It picks better than random, and yes, it plays Doom.

Later I found that privatemode.ai had arrived at a similar approach independently. Science and progress have always relied on access to information.

Also in this issue: a small fable about the first rule of cloud-native. One slow S3 write froze my entire blog, and it was a reminder that a system is only as strong as its weakest link. It's worse when the whole thing is just one link.

The code is on github under dooreelko/jeb

#AI #LLM #MachineLearning #OpenSource #llamacpp #CloudNative #AWS #Serverless #SoftwareArchitecture #Reliability

#ai #llm #machinelearning #opensource #llamacpp #cloudnative

Sascha Fedorenko 🇨🇭🇺🇦 @[email protected] · 2026-09-29 · 10:00 UTC

https://the-well-architected-cloud.com/blog/do-androids-dream-of-flappy-birds/

TypeSafe released Jev, a "decision model" that's "just like an LLM, but not an LLM". They didn't share how it works.

So I guessed. My hypothesis: it's a regular LLM that stops before generating text and reports how likely each option is to continue the prompt.

To test it, I built Jeb: an unmodified LLM, llama.cpp and the raw next-token probabilities. It works. It picks better than random, and yes, it plays Doom.

Later I found that privatemode.ai had arrived at a similar approach independently. Science and progress have always relied on access to information.

Also in this issue: a small fable about the first rule of cloud-native. One slow S3 write froze my entire blog, and it was a reminder that a system is only as strong as its weakest link. It's worse when the whole thing is just one link.

The code is on github under dooreelko/jeb

#AI #LLM #MachineLearning #OpenSource #llamacpp #CloudNative #AWS #Serverless #SoftwareArchitecture #Reliability

#ai #llm #machinelearning #opensource #llamacpp #cloudnative

Sascha Fedorenko 🇨🇭🇺🇦 @sascha_fedorenko · 2026-09-29 · 10:00 UTC

https://the-well-architected-cloud.com/blog/do-androids-dream-of-flappy-birds/

TypeSafe released Jev, a "decision model" that's "just like an LLM, but not an LLM". They didn't share how it works.

So I guessed. My hypothesis: it's a regular LLM that stops before generating text and reports how likely each option is to continue the prompt.

To test it, I built Jeb: an unmodified LLM, llama.cpp and the raw next-token probabilities. It works. It picks better than random, and yes, it plays Doom.

Later I found that privatemode.ai had arrived at a similar approach independently. Science and progress have always relied on access to information.

Also in this issue: a small fable about the first rule of cloud-native. One slow S3 write froze my entire blog, and it was a reminder that a system is only as strong as its weakest link. It's worse when the whole thing is just one link.

The code is on github under dooreelko/jeb

#AI #LLM #MachineLearning #OpenSource #llamacpp #CloudNative #AWS #Serverless #SoftwareArchitecture #Reliability

#ai #llm #machinelearning #opensource #llamacpp #cloudnative

Sascha Fedorenko 🇨🇭🇺🇦 @[email protected] · 2026-09-29 · 10:00 UTC

https://the-well-architected-cloud.com/blog/do-androids-dream-of-flappy-birds/

TypeSafe released Jev, a "decision model" that's "just like an LLM, but not an LLM". They didn't share how it works.

So I guessed. My hypothesis: it's a regular LLM that stops before generating text and reports how likely each option is to continue the prompt.

To test it, I built Jeb: an unmodified LLM, llama.cpp and the raw next-token probabilities. It works. It picks better than random, and yes, it plays Doom.

Later I found that privatemode.ai had arrived at a similar approach independently. Science and progress have always relied on access to information.

Also in this issue: a small fable about the first rule of cloud-native. One slow S3 write froze my entire blog, and it was a reminder that a system is only as strong as its weakest link. It's worse when the whole thing is just one link.

The code is on github under dooreelko/jeb

#AI #LLM #MachineLearning #OpenSource #llamacpp #CloudNative #AWS #Serverless #SoftwareArchitecture #Reliability

#reliability #softwarearchitecture #serverless #aws #cloudnative #llamacpp

Sascha Fedorenko 🇨🇭🇺🇦 @[email protected] · 2026-09-29 · 10:00 UTC

https://the-well-architected-cloud.com/blog/do-androids-dream-of-flappy-birds/

TypeSafe released Jev, a "decision model" that's "just like an LLM, but not an LLM". They didn't share how it works.

So I guessed. My hypothesis: it's a regular LLM that stops before generating text and reports how likely each option is to continue the prompt.

To test it, I built Jeb: an unmodified LLM, llama.cpp and the raw next-token probabilities. It works. It picks better than random, and yes, it plays Doom.

Later I found that privatemode.ai had arrived at a similar approach independently. Science and progress have always relied on access to information.

Also in this issue: a small fable about the first rule of cloud-native. One slow S3 write froze my entire blog, and it was a reminder that a system is only as strong as its weakest link. It's worse when the whole thing is just one link.

The code is on github under dooreelko/jeb

#AI #LLM #MachineLearning #OpenSource #llamacpp #CloudNative #AWS #Serverless #SoftwareArchitecture #Reliability