Sam Saffron

Cofounder and co-CEO , rubiest, SQL fan, open source developer.

[Sam Saffron @samsaffron]

@nateberkopec @fleetingbits At @discourse we started systematic full codebase scanning about once a week and have a dashboard that is always showing us the top issues with candidate patches that pass CI. An abliterated DeepSeek would probably do better than Sol 6.1 which is the best at this atm.

[Sam Saffron @samsaffron]

Discourse now allows you to set up "chat first" style communities out of the box: https://t.co/6KeIvT1mPu

[Sam Saffron @samsaffron]

It is working!

[Sam Saffron @samsaffron]

If you have @typesafeai Jev or any decision model configured in term-llm, you will get smart reranking of ag results starting today: https://t.co/L7djYq9OZB

[Sam Saffron @samsaffron]

If you have @typesafeai Jev or any decision model configured in term-llm, you will get smart reranking of ag results starting today: https://t.co/L7djYq9OZB

[Sam Saffron @samsaffron]

@solomonstre you mean ama?

[Sam Saffron @samsaffron]

This is looking very promising: use @typesafeai Jev as a reranker to improve grep/rg results for term-llm

[Sam Saffron @samsaffron]

I have thought about this as well, maybe the harness of the future is just 30 paragraphs of text and model bootstraps itself. Models are way too jagged today to make this a reality but I am not ruling it out. For now though I think detection and auto repair is probably a better approach.

[Sam Saffron @samsaffron]

Writing a agent harness is easy except for the 98737 edge cases.

[Sam Saffron @samsaffron]

@mani_koppala Yeah, if this is a direction you are headed you need to be in it for the long haul. My Jarvis (term-llm bot) has been going really well for the past 6 months plus, but there are tremendous amounts of details getting this to work well.

[Sam Saffron @samsaffron]

Oh my: https://t.co/wkeUgCZ5kl

[Sam Saffron @samsaffron]

Our designer and engineer Kris came up with a wonderful use case for decision models in search, love it. https://t.co/Aqei2PO6FB

[Sam Saffron @samsaffron]

@anaisbetts It has a gazzilion tools, very easy for a non fine tuned llm to get confused, I did a test with term in the link, no magic in dsh, solid harness but not magical

[Sam Saffron @samsaffron]

No, this is wrong, there is no magic voodoo "feels" in AI harnesses, we all do the same things, tools, system prompts, skills, etc... https://t.co/WC3jTnoZd6

[Beamsters @Beamsters1]

Deepseek 4.1 Flash 1017 needs Deepseek Harness to perform well.

Terminal 4 bench 32 without and 40+ with dsh.

[Sam Saffron @samsaffron]

@nateberkopec Same, and for some reason OpenAI has been impossible for me @reach_vb my account is in some sort of verification limbo sent my passport over, next up is a pint of blood. Given up on daybreak, some other team members have it so it is fine.

[Sam Saffron @samsaffron]

22 GPU hours a month on your existing $20 dollar Google AI plan to run Qwen 3.8 Flash Next on a RTX PRO 6000, ~45 tok/s 10k/s prefill.

https://t.co/Ba2XlsR2X5

[Sam Saffron @samsaffron]

@hrkrshnn I can take the same tools and same system prompt and run on my harness.... there is no magical and mystical properly here.

[Sam Saffron @samsaffron]

On the weekend I was brainstorming how to make term-llm easier to work on for agents ... today I merged this monster https://t.co/m8JsVmnv2e

[Sam Saffron @samsaffron]

@joshmo_dev @AskVenice Be sure to try out term llm has super solid Venice support (model listing, images, audio, inference)

[Sam Saffron @samsaffron]

If you are using @AskVenice with Anthropic models, be VERY careful to use a harness that adds venice specific caching instructions on the open ai completion API. Works in term-llm now, but most harnesses will miss this nuance cause it is very unexpected.

[Sam Saffron @samsaffron]

I am about to learn how much 7% of a $200 dollar Codex plan gets me in Luna High tokens... I bet this is close to infinity.

[Sam Saffron @samsaffron]

👋Argon, yeah I agree with you, you came up with a simpler fix this time.