
we recently released FrontierMemory (frontiermemory.com/), a new way to evaluate dreaming, and are spending a lot of time improving how agents dream. if that sounds interesting, we're hiring researchers!

agent artist @letta_ai | @LTIatCMU @bitspilaniindia

we recently released FrontierMemory (frontiermemory.com/), a new way to evaluate dreaming, and are spending a lot of time improving how agents dream. if that sounds interesting, we're hiring researchers!

sleep-time compute was a big reason I joined @Letta_AI's research team last year. very cool to see it evolve into "dreaming" and become a key primitive for memory!
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. anthropic.com/research/compo…

FrontierMemory is now on @FireworksAI_HQ's Specialized Intelligence Index. Picking a model for a long-running agent? You can now compare how well each one learns from its own past attempts, combined with cost. View FrontierMemory results under RSI: fireworks.ai/specialized-int…
This tweet is unavailable

EVALUESTEER: Measuring Reward Model Steerability Towards Values and Preferences 🎓 Led by @GhateKshitish 🎤 Presented by me! 📍 Hall H, Poster Session E | #2197 🗓️ Oct 26, 4:45–6:15 PM
🚨New paper: Reward Models (RMs) are used to align LLMs, but can they be steered toward user-specific value/style preferences? With EVALUESTEER, we find even the best RMs we tested exhibit their own value/style biases, and are unable to align with a user >25% of the time. 🧵


nvidia acquisition soon? 👀
1/ muse spark open is the most capable US open-weight model to date. 816B total params, 42B active. 1M-token context. natively multimodal. built for coding and agentic work like long-horizon tasks, hard software engineering, multi-step reasoning. america! 🇺🇲

sota on chonkbench 🇫🇷
Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.

Today we’re announcing Not Diamond Code, the world’s most powerful intelligent model router for long-horizon coding agents. Not Diamond works with any gateway or harness, including Claude Code, to select the best model and reasoning effort for each step, reducing costs by 20-65% without impacting quality.



> maybe continual learning should just focus on loading the right information up for icl (ie into the context window)
13/ What I take away from all this is that all the training to create the model is engineered around crafting the best possible ICL mechanism. Further training degrades this mechanism, at least for knowledge acquisition, and maybe continual learning should just focus on loading the right information up for icl (ie into the context window) So this provides an answer to the architecture question, at least for now: the channel worth engineering for memory is the one that comes with addresses ie the context. Paper: arxiv.org/abs/2607.11020 Code and data to follow shortly Done at @baseten.

Really nice research from @baseten on parametric learning vs in-context learning. ICL is the 'bitter lesson' of continual learning. Each new generation of model will be better at ICL than the last. Mechanisms for CL through context will ride the waves of model improvements.
13/ What I take away from all this is that all the training to create the model is engineered around crafting the best possible ICL mechanism. Further training degrades this mechanism, at least for knowledge acquisition, and maybe continual learning should just focus on loading the right information up for icl (ie into the context window) So this provides an answer to the architecture question, at least for now: the channel worth engineering for memory is the one that comes with addresses ie the context. Paper: arxiv.org/abs/2607.11020 Code and data to follow shortly Done at @baseten.

Agents can now not only rewrite their memory, but also rewrite the execution code they run in (the harness) through Mods. Mods (inspired by Pi's extension system) allow for harness-level changes like: - modifying context - injecting custom tools - customizing the statusline

i like my training like i like my agents: memory-native 💤
Letta @Letta_AI Jun 25 Article How can we build agents that learn like people? tl;dr - Agents that learn like people (through experience, teaching, & self-study) will be powered by memory models: models specialized in creating and curating token-space memory that transfers



Was using Fable 5 to write my self-improving agent harness. Anthropic flagged it as frontier AI research. The steering vector kicked in and it started implementing hermes agent 🤨
Was using Fable 5 to write my world model training code. Anthropic flagged it as frontier AI research. The steering vector kicked in and it started implementing JEPA 🤨