Nobody became a developer to write documentation. But the docs still need to get written — PRDs, README updates, architecture decisions, onboarding guides.

Wispr Flow lets you talk through it instead. Speak naturally about what the code does, how it works, and why you built it that way. Flow formats everything into clean, professional text you can paste into Notion, Confluence, or GitHub.

Used by engineering teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works system-wide on Mac, Windows, and iPhone.

Try Wispr Flow free

Lourie et al. [FAIR at MSL Meta, New York University]

♥ 575 LLM Scaling

To save costs, we use scaling laws to predict a large model's performance using tiny, inexpensive test runs. However, these predictions break down at small scales, which creates a gap where insights from small-scale experiments don’t translate into massive production systems.

Small models are sensitive to hyperparameter settings and any suboptimal choice injects heavy noise that completely hides the underlying scaling law. Uncovering these hidden laws requires searching through hundreds of configuration variations.

When models grow in scale, their tuning undergoes a geometric transformation where its dimensions drop in number and this causes near-optimal settings to fill a much larger fraction of the space. At the same time, smaller models demand extensive searching to tune properly, whereas large models naturally become far more forgiving.

By conducting thorough hyperparameter searches on cheap, single-GPU models, researchers can establish accurate performance trends and confidently evaluate fundamental architectural choices, such as where to place structural normalization layers, at a fraction of the traditional cost.

Wang et al. [Johns Hopkins University, Princeton University, Microsoft]

♥ 776 Transformers bycloud’s pick

LLMs compute a rich multidimensional snapshot of their thoughts at its deepest layer, it discards that state and passes only the selected word back to the bottom layer. This narrow feedback loop forces systems to either write out verbose, explicit trails of step-by-step logic or lose track of intricate ideas.

To widen this channel, researchers created a full-bandwidth transformer that keeps the internal conversation flowing between steps. They used a technique called latent feedback decoding, where the system blends its previous top-layer hidden state directly with the newly selected word, then feeds that combined signal back into the starting layer.

This creates a continuous internal state where ideas, strategic plans, and partial calculations re-enter the model with a fresh depth budget. This approach preserves standard model architectures and adds one percent computational cost.

Full-bandwidth models can match or approach standard systems trained on up to fifty percent more data across mathematical, coding, and general language benchmarks.

Additionally, because these models can digest complex reasoning inside their continuous feedback loop instead of narrating every detail, they solve tough problems with higher accuracy while producing noticeably shorter, more direct written answers.

Panfilov et al. [MATS Research, ELLIS Institute Tübingen, Tübingen AI Center, AI Sequrity Company, Max Planck Institute for Intelligent Systems, Snyk University of Tübingen]

♥ 13k LLM Reasoning Signatures

Most AI providers hide their models' step-by-step internal reasoning inside encrypted code blocks. However, these encrypted "thinking" blocks are fully compatible across different models, user accounts, and sessions.

This cross-compatibility creates a vulnerability, which allows anyone to take an encrypted trace from a top-tier model and pass it to a cheaper, less-guarded model from the same vendor, which willingly translates the hidden text into plain language.

This simple replay technique turns smaller models into unexpected decryption tools. Because safety guardrails are often weaker on lower-cost models, they freely transcribe the precise problem-solving steps and sensitive logic that the primary model was programmed to protect.

When researchers tested this approach on publicly shared online session logs, they successfully recovered hundreds of sensitive items, such as active software keys, passwords, and personal information, that users had unknowingly exposed within the encrypted code. The flaw also makes it possible to quietly bypass safety filters or steal competitive AI reasoning at scale.

Zhu et al. [Chinese Academy of Sciences, School of Artificial Intelligence, Tsinghua University]

♥ 500 LLM RL

When we train LLMs via supervised fine-tuning method, their capabilities suffer from task conflicts when trained sequentially across multiple capabilities. Teaching a model a new skill drastically degrades its performance on previously learned ones.

RL avoids this problem by allowing multiple complex capabilities to stably coexist and continuously build upon each other. When you analyze the updates inside the model, you would see that RL acts like a precision instrument rather than a broad brush.

Instead of making sweeping adjustments across the system, RL induces remarkably tiny, sparse parameter changes. The directions of these updates for different tasks are almost completely perpendicular, or orthogonal, to one another. This geometric independence means that tuning a model to write code leaves its mathematical reasoning pathways virtually untouched.

Using this insight, this paper introduced Parallel-RL, a framework where distinct capabilities are trained independently in parallel and then seamlessly merged together without performance loss.

This approach allowed researchers to build a far more efficient, flexible, and scalable mode by ensuring models can continuously acquire specialized skills without compromising their broader intelligence.