SATURDAY · SEPTEMBER 12, 2026 · ISSUE 12 · YESTERDAY No. 12
Daily News — 2026-09-12
~6,400 packages released on PyPI·26 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
Latent concept prediction halves the tokens needed to match a strong 7B baseline — a sample-efficiency result worth taking seriously.
The central claim here is that next-token prediction is leaving performance on the table, and this architecture makes a concrete case for why. NCP-ArchPreview trains an 8.9B-parameter model on the Dolma-3 dataset — just…
A working desktop-wide DLSS 5 wrapper built on a leaked runtime — technically interesting, legally unambiguous, and honest about both.
NeuralScreen wraps NVIDIA's DLSS 5 neural renderer around your entire Windows desktop — not just games that have been patched to support it, but every window, video, and photo on screen at once. The mechanism is a…
Papers
Latent concept prediction halves the tokens needed to match a strong 7B baseline — a sample-efficiency result worth taking seriously.
NCP-ArchPreview jointly trains next-token and next-concept prediction, reporting gains in both pretraining efficiency and downstream performance.
Synthetic block puzzles teach spatial reasoning that transfers to real scenes better than real-scene annotation does.
A study showing LVLMs trained on synthetic block-stacking tasks improve 3D spatial reasoning and transfer to real-world visual tasks.
A practical, honestly scoped workflow that treats frequency-domain artifact type as the decision variable, not just the score to minimize.
Mi-Ripple targets the lattice artifacts that accumulate across repeated AI image edits, using spectral filtering to separate and remove them from texture.
Cross-scale distillation from a 1.7B teacher is what separates a recovered pruned encoder from one that collapses—self-distillation fails outright under the same recipe.
X-AuT progressively prunes audio-encoder layers in speech LLMs and restores accuracy via cross-scale distillation and LoRA adaptation.
Drops every flow-specific architectural prior and still tops Sintel, KITTI, and Spring — the simplicity is the argument.
A hierarchical transformer for optical flow estimation that removes task-specific inductive biases using window, shifted-window, and global attention.
Keeping the noisiest 5% of KV positions in full precision while fusing the rest into a 4-bit kernel is a small structural insight with measurable decode-stage payoff.
HyQuant quantizes LLM attention at low bit-width while keeping vertical-line tokens and local windows in high precision to preserve accuracy.
9 more tool picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
I2T loss, not reconstruction FID, is the more honest cross-tokenizer signal for unified multimodal training—and the image token space quietly shapes text modeling too.
Controlled autoregressive testbed uses per-task validation losses to isolate how image tokenizer design shapes unified text-image model performance.
Recursive global–local–global construction, not a better representation, is what closes the gap between flat scene reconstruction and fine-scale fidelity.
RCWM builds complex 3D worlds as executable code from a single image using recursive global and local scene reconstruction.
Reframes world-model control as an evidence-routing problem, achieving top camera accuracy on DAVIS and OpenVid-1M without updating a single pretrained weight.
A training-free interface for camera and time control in frozen autoregressive video world models via correspondence-guided self-attention routing.
Reasoning improves by diverging from self-generated flawed traces, not by imitating privileged correct ones - and the reflection behavior survives.
A training method that improves LLM reasoning by pushing models away from their own flawed outputs via a dynamic gating mechanism.
Tools & packages
A working desktop-wide DLSS 5 wrapper built on a leaked runtime — technically interesting, legally unambiguous, and honest about both.
A Windows desktop tool that applies DLSS 5 neural rendering in real time, with user presets, audio recording, and one-window mode on RTX 30/40/50 GPUs.
A real fly connectome wired to a real trading account — the honest admission that profitable learning hasn't been shown is what makes it credible.
A fly-connectome simulation with experimental memory and guarded Coinbase AgentKit trading actions.
A prompt-based Copilot configuration that enforces predict-first, hint-second discipline — the learning mechanism is the whole product.
Reframes AI agents as teachers rather than code-dumping machines, prioritizing explanation over raw output.
A local proxy that makes Copilot's model identity a runtime variable rather than a config file value — the architecture is sound and the friction reduction is genuine.
A local proxy for VSCode Copilot Chat that switches AI models dynamically without reloading.
A focused, well-documented TopK kernel that trades generality for genuine memory-bandwidth gains in the exact shapes sparse attention and sampling actually use.
Optimized TopK CUDA kernels for DeepSeek Sparse Attention routing and token sampling.
A live-updating personal knowledge skill whose real contribution is the architecture: thin loader, version-controlled markdown, daily automation keeping opinions current.
Packages a principal engineer's experience, tools, and workflows so you can think and build at that level.
10 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
Gives the model a live JavaScript runtime and raw CDP instead of a fixed action vocabulary — powerful, but the open execution environment demands careful isolation.
Minimal TypeScript web agent, iteratively optimized against live browser benchmarks.
A shared discrete action vocabulary lets one VLM drive four robot platforms — the interface, not the model, is doing the embodiment work.
Show-Harness argues a stock VLM agent alone is enough to control robots, no specialized robotics framework needed.
A concrete agent loop for live game modification that's more honest about its prototype status than most repos at this stage.
SuperAstra lets you modify SNES games with AI in real time as you play them.