Hello, author here excited to discuss with everyone. so the context is More agents are running for longer turns and at greater context length. This is A * T * C ~= Cubic problem. Unaddressed by SOTA approaches. "fak" Fused agent kernel solves this turning O^3 into O(n). So far proven on small to middle sized tasks at ~4x SOTA. Unlike most other approaches there is no real downside or effort required to get the benefit. While its' still early in terms of proving it out on large scale and hyperscaler workloads these early results show promise. Thank you
Sharing Zen: runs multiple Code CLI instances for peaceful parallel task execution.
Zen allows you to:
+ Run multiple headless Claude Code CLI instances simultaneously.
+ Calm unified results (status, time, token usage)
+ Relax "5-hour limit reached" lockout fears with easy token budget limits
+ Get more value out of your Claude MAX subscription with scheduling features. (--run-at "2am")
+ Learn more about how Claude Code uses tools and other inner workings
Control usage and budget for groups of work or per command
Hey folks! We are really excited to release this. It's been years in the making and it's very much the full application. We make $ from selling the enterprise supported version and are working on new paid enterprise hub product. I personally really see this Open Core release as version 1 of many versions to come over the years. If you have any questions AMA :)
I'm curious why less features is a goal here?
I wish VueJS had more features, or at the very least more tooling. Vue has progressed a lot but there is so much more room https://github.com/vuejs/vue-devtools
Hey HN! We are Anthony and Pablo and we’re building software for human supervision of AI data.
Diffgram started a little over 2 years ago. Recently, as a tiny self funded team, we have won a few customers away from larger venture backed firms that have raised over 40M. While there are gaps, as an underdog we have been able to keep pace with and even exceed these 10x+ larger teams.
Our focus is the basics, with a core assumption that the AI models are one of the best forms of speedup. We have also done our best to move iteratively. One example https://bit.ly/38OxZ7T A visual of internal commit history: https://bit.ly/34RDCRN
The big idea is ongoing human supervision is crucial, important, and cost effective for a large class of practical AI applications. Diffgram is scoped to cover Training Data between Raw Data and Modeling. Diffgram covers Data Prep, Tasks, Image & Video Interface, Data Management, etc.
Teams often get a baseline case covered, for example spatial location (drawing a box, or pixel by pixel etc). The problem is to add depth and breadth. Long story short, representing truly useful, editable, training data is hard.
In Diffgram you can solve this by representing as much of this manual, unspoken work, in reusable abstractions. Many of this happens by default - there’s no “new language” to learn - it’s automatic organization of the stuff you already know.
Benefits 1) Supervision productivity improvements 2) Data engineering productivity: Less up front (integrations), less maintenance work (wheel and spoke included), less risk (central schema definition). 3) Project: Higher chance of overall product success: See https://bit.ly/3n2gfLB
Hey HN! We are Anthony and Pablo and we’re building software for human supervision of AI data.
Diffgram started a little over 2 years ago. Recently, as a tiny self funded team, we have won a few customers away from larger venture backed firms that have raised over 40M. While there are gaps, as an underdog we have been able to keep pace with and even exceed these 10x+ larger teams.
Our focus is the basics, with a core assumption that the AI models are one of the best forms of speedup. We have also done our best to move iteratively. One example https://bit.ly/38OxZ7T A visual of internal commit history: https://bit.ly/34RDCRN
The big idea is ongoing human supervision is crucial, important, and cost effective for a large class of practical AI applications. Diffgram is scoped to cover Training Data between Raw Data and Modeling. Diffgram covers Data Prep, Tasks, Image & Video Interface, Data Management, etc.
Teams often get a baseline case covered, for example spatial location (drawing a box, or pixel by pixel etc). The problem is to add depth and breadth. Long story short, representing truly useful, editable, training data is hard.
In Diffgram you can solve this by representing as much of this manual, unspoken work, in reusable abstractions. Many of this happens by default - there’s no “new language” to learn - it’s automatic organization of the stuff you already know.
Benefits
1) Supervision productivity improvements 2) Data engineering productivity: Less up front (integrations), less maintenance work (wheel and spoke included), less risk (central schema definition). 3) Project: Higher chance of overall product success: See https://bit.ly/3n2gfLB
Scale AI is focused on the outsourcing side. We have an integration with them. I think the future of this is so much more then outsourced labor though. Subject matter experts and people in existing roles are great fit for supervising these systems.
Re: 'Time limits are detrimental and discriminatory' the short answer is that it's really testing if you already know the answer - some of these original algorithms took decades to discovery the first time. My interview process if a bit different but still very tough. I talk about some of my opinions on that here https://medium.com/@anthony_sarkis/software-engineering-path...