gemma4-31b can score 31% or 89% on the same questions. The only difference is the harness, and the code ships so you can regenerate every number on a CPU.
The latest Krea 2 Turbo distillation checkpoint (chk26K) drops generation to four steps while shaving 46% off the error rate versus the eight-step Turbo teacher. That means near-two-second 2K images on consumer GPUs with no loss in detail. It's a clear signal that step-distillation, not bigger models, is where the next speedups in image diffusion are coming from.
HAWKEYE grounds coding agents in a minimal taxonomy: one unit test per optimization strategy per architecture, about ten expert-written tests per new accelerator. Scaled with test-time compute, agents then write hardware-aware kernels that match or beat hand-tuned PyTorch and Triton across Ampere, Hopper, Blackwell and MI350, in FP8, NVFP4 and MXFP4. It targets the real bottleneck of the ML-chip explosion: software support, not hardware.
A fragility grid running 12 models over 3,679 multiple-choice items under 26 defensible harness setups finds score is a band, not a point: gemma4-31b swings 31 to 89 percent on configuration alone, and 4 of 12 models can reach rank one under some config. Config-fragile items carry 95.7% of a model pair's score gap and correlate with the discrimination that benchmark-compression preserves. Option order matters less than the scoring axis, text likelihood versus sampled tokens, so leaderboard order is partly manufactured. All per-item records and analysis code ship so every number regenerates in seconds on a CPU.
Reusable solver archives are becoming standard training data for neural PDE operators, so researchers probed what happens when that data is tampered with. They introduce cross-parameter relinking, a poisoning primitive that makes a triggered input trigger a valid but physically wrong solution from the same PDE family. Across 476 campaigns, Fourier Neural Operators hit a 1.0000 backdoor success rate on advection-diffusion and 2D Navier-Stokes while keeping clean error low. Clean-error checks and smoothness tests don't catch it, exposing a structural validation gap.
METR scraped every public discovery time series it could find and looked for slope changes around a January 2026 breakpoint. Vulnerabilities jumped dramatically across cURL, OpenSSL, Firefox, Microsoft, and aggregate databases. Math accelerated somewhat, but harder to measure; optimizations barely moved. The punchline: exploitation databases (KEV) grew far slower than known-vulnerability counts, and not all that matters is public, since labs keep internal discoveries quiet.
Standard training pools freeze the goal distribution as a model scales. SPADE instead splits one LLM into an Environment Designer that writes executable, Gym-style training worlds in code and a Reasoning Agent that learns in them, with a regret signal that forces environments to track the edge of the learner's ability. At 30B parameters it beats the strongest fixed-environment baseline by +5.3 across eight held-out benchmarks and by +5.7 to +13.9 on tool-use tasks, and the margin grows with scale. It is a concrete step from curated data toward open-ended, recursive self-improvement.
A tinkerer simulated low-Earth-orbit radiation by randomly flipping bits inside Qwen2.5-Coder-3B's weights in memory. Models degrade shockingly fast: after a handful of bit-flips, the coder's human-eval score collapses from 85% toward zero, and a single flipped exponent bit can wreck a 49-billion-bit model. The results argue that plans for spaceborne or radiation-exposed data centers need ECC and fault handling as a serious design constraint, not an afterthought.
A team improved the long-contested matrix multiplication exponent to ω < 2.371177, shaving the prior best bound of 2.371339. The trick: reframe the laser-method optimization problem, then let a bespoke ML optimizer refined by AlphaEvolve hunt for the answer. Small on paper, but it is fresh proof that AI now steers frontier math, not just verifies it.
Investigators recovered a Russian Molniya drone carrying an unencrypted Nvidia Jetson Orin chip that navigated, chose a target, and detonated with no radio link to any human pilot, reportedly the first documented civilian deaths from a fully autonomous strike. It failed to clear an apartment building, hit a wall, and killed three people sheltering nearby. This goes beyond the last-mile autonomy both sides already use.
OpenAI banned a cluster of Russia-linked ChatGPT accounts behind an unusually elaborate covert influence campaign: a fake Israel-based "International Burke Institute" padded with copied and misattributed academic work, plus a "sovereignty" index flattering Moscow and criticizing the West. Operators prompted in Russian, ordered ChatGPT to erase any Russian linguistic traces, and used VPNs to dodge the Russia access ban. Reach was small, but the setup marked a shift: classic influence tradecraft, now run with AI at its core.
Alibaba's MiniMax-H3 Fun ControlNet-Union packs five conditions (Canny, depth, HED, MLSD, pose) plus video inpainting into a single checkpoint, so creators stop swapping weights per video-to-video job. Guidance-distilled, it runs one forward pass per step with zero classifier-free guidance. It hooks into only 5 of H3's 50 transformer layers via zero-gated projections, keeping control cheap without nuking quality.
NVIDIA ships a disassembler but no SASS assembler for modern data-center GPUs, which blocks controlled binary rewriting. F2Asm learns exact 128-bit instruction encoders as affine maps over F2 via Gaussian elimination, trained on 3,225 real CUBINs from Hopper, Blackwell, and Rubin. In round-trip tests every executable text section matches the original exactly. It's reportedly the first open-source SASS assembler to support Rubin SM107.
You’re caught up.
That was all 12 stories in the 25 August issue.