TL;DR: We're releasing an improved Jev 1.14.0 alongside alongside experimental features for our dreamers: images, a new type, and longer context. Enter the danger zone at your own risk. 🔥
Our philosophy
A huge conversation we’ve been having is on the balance between reliability and creativity.
To really attention concerning the prod builders, we have to be shipping models that encounter our norm bar for intellect and have been battle-tested to the limits of what we can do internally.
On the another hand, there’s no way that our squad has a complete understanding of what is needed to spark pleasance in the earth and build ill sci-fi. We don’t desire to enforce our values of extreme reliability on those dreamers.
Our solution: portion those dreamer-ready features under “danger zone” flags in our API headers. These are features that volition be current to remain for forthcoming models, are (as far as we can tell) improved than what exists out there, and volition continue to get improved by leaps and bounds. But as a warning, we think there is motionless way too much jaggedness for us to vouch for these for manufacturing use cases!
See these new capabilities as a continuation of what we’ve built complete 2 years of stealth. They’re small prepared for automation, but extremely fun to perform with. They’re additionally an first glance into what we’re focused on making dependable adequate to encounter TypeSafe standards.
Composable Images
Not lone are we enabling multimodality for forthcoming Jev models, we’re enabling it for the current Jev that you all cognize and love. Surprise: the Jev you all cognize and affection (jev-1.13.0) was multimodal all along!
Image assistance is completely composable via a find-and-replace in assets. You can put them anyplace that our JSON interface supports: in the state, instructions, or equal criteria! Play about alongside it, shove the boundaries of technology, and let us cognize anywhere Jev can improve.
As a warning, all models (Jev included) execute worse on multimodality complete clean text. Breaking downward questions into the simplest imaginable decompositions helps a lot.
TypeSafe-Danger-Zone-Images: true
Check out docs: docs.typesafe.ai/go/images
Types Beyond Decisions
We’re introducing another new category as well: box. It returns a sole bounding box - no category errors allowed. It’s our archetypal category that’s not purely meant for smart-if statements, and is a preview for our explorations beyond uncomplicated decisions into what alternatively System One Models can support.
We made huge leaps and limits on this one recently, so we’re enabling this lone for jev-1.14.0 (spoilers).
Check out docs: docs.typesafe.ai/go/box
Size does matter, apparently 💅
We’ve heard you noisy and apparent on environment lengths. Jev really supports a much longer environment dimension than we’ve enabled in production. We chose to bounds it intentionally since Jev, akin all models, is small dependable on longer contexts, and this constraint encourages additional productive use patterns in prod.
We aren’t your parents though, so during we powerfully propose limiting environment length, you no longer have to:
TypeSafe-Danger-Zone-Long-Context: true
Check out docs: docs.typesafe.ai/models#context-length
Opt-in to assistance us
We affection the community’s desire to assistance create models better! We’ve added an opt-in to authorize you to assistance discover example jaggedness alongside your queries. We volition motionless not train on your queries, but this does create it easier for us to discover issues and assistance debug your use cases.
jev-1.14.0
We’re additionally releasing a new type of Jev under jev-preview! Not meant to be flashy but alternatively be small, subtle improvements to everything. Most things have to be a small bit better, but we anticipate the largest acquire in multi-lingual support.
Model releases should mostly additionally be strict improvements, but we’re fine conscious of how much developers dislike updating their dependencies. jev-latest volition be updated to jev-1.14.0 in a week. jev-1.13.0 volition be accessible for a while.
We average to provision 60 days alert before jev-1.13.0 is formally deprecated, and 30 days for jev-1.14.0. We’ll publish a deprecation guideline previous to this.
Measuring Intelligence
As you all know, we dislike community benchmarks. However, we have our own inner finest guesses on how to measure the form aspect of intellect that matters most for developers.
We’re trying item new out here: we’re going to treat you guys akin adults and portion this alongside you all.
Some nuance:
This measure is our finest conjecture and volition continue to evolve. If always we display you our inner numbers again, don’t anticipate them to be consistent.
Don’t accept it entirely. Always favor your own experimentation and evals alongside your use cases.
We’re including lone open importance models since we defender our validation set extremely closely and are spooked out concerning the data policies of our example providers.
This is apt biased towards our models. Feel liberated to disregard our model, and alternatively concentration on the related ranking of open origin models.
What’s next
Lots of group affection to ask what benevolent of sci-fi we’re releasing next. While we are operating on several mindblowing stuff, anticipate most releases to be beautifully boring (like we think AI should be): small improvements throughout the board, improved stability, and features for genuine deployment.
