AI FRONTIER
Signals worth understanding before they become obvious.
DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency
RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning using Miles on long-horizon agent tasks with sessions up to 1M tokens, using just 1,200 curated tasks in 18 days on 144 H200 GPUs.
- Terminal-Bench 4.1 pass@1 improves from 5.6% to 12.0%.
- Checkpoints are available under Apache 2.0 in BF16, FP8, and NVFP4.
- The model retains Qwen3.8-27B's architecture, tokenizer, chat template, image understanding, and 256K-token context window.
- Training ran for weeks on 144 H200s, 63,000 GPU-hours in total.
For agent builders, DaisyRose-27B shows that roughly 1,200 curated tasks and 18 days of training can double a terminal benchmark score while shipping Apache 2.0 checkpoints.
How OpenAI’s Performance Factory keeps ChatGPT fast
OpenAI built the Performance Factory, a system of Codex-powered workflows that finds and fixes performance issues to help keep ChatGPT fast.
- The Performance Factory is a system of Codex-powered workflows.
- It finds and fixes performance issues.
- OpenAI states the goal is to help keep ChatGPT fast.
For teams building AI products, this shows how code-generation workflows can automate performance triage and fixes instead of relying on manual tuning alone.
A Differential Theory of Quality-Resource Trade-offs for AI Model Selection: The Multiresource Frontier and Its Minimax Price
The paper develops a differential theory representing quality-resource frontiers as a price-independent plane-curve relating the effective number of active resource channels, N_eff = s²/κ, to the remaining log-error depth, ΔS = ln(ε/ε∞). When each of N resources contributes one active channel, the theory gives the para...
- The paper reports evaluating the theory on 117 published frontiers.
- The theory derives minimax resource prices λ*_i = α_i/(N+1) with an exact worst-case regret of ln(N+1) nats.
- When each of N resources contributes one active channel, the theory gives the parameter-free operating rule N_eff = 1 in delivered units.
- The theory is implemented in the open-source package nn-rank, which derives prices from measured quality and cost, ranks candidates, and reports supported choices and minimax regret.
AI builders can use nn-rank to derive resource prices from measured quality and cost and rank candidates without hand-chosen exchange rates, while obtaining a quantified minimax regret guarantee.
Bits, Not Boxes: Label-Free Layer Selection for Visual Localization in VLMs
The paper introduces WhereBits, which turns where-to-look bits measured on unlabeled questions into a layer prior and combines it with per-head contrastive scores to select layers that localize the queried object, without box annotations or fixed rules.
- The paper reports WhereBits reaches 0.72 COCO pointing accuracy, against 0.22 for mean attention.
- Across eight VLMs, WhereBits matches the pointing accuracy of box-tuned layers on COCO and RefCOCO and ranks target tokens better.
- The advantage holds on a held-out model family and at higher resolutions.
- In training-free zoom-in, its crops raise V* accuracy by 6.0 points under forced-choice evaluation.
For builders of VLM localization pipelines, layer selection can be recalibrated for a new model or resolution using only unlabeled questions, removing the need for box annotations.
Tencent releases Hy4-flash: a 220B-parameter, 12B-active multimodal MoE model
Tencent's Hy4-flash uses a Mixture-of-Experts architecture with 220B total parameters and 12B active per token, natively accepts text and image inputs, supports a context window of up to 1M tokens, and is released under the Tencent Hy License.
- 220B total parameters (including 64B Engram parameters) with 12B active parameters per token; the 42-layer language backbone has one dense layer and 41 MoE layers.
- Context length of 1M (1,048,576 tokens); the vision encoder is adapted from SigLIP2-So400m-Patch16-Res512 and connects to the backbone through an MLP projector.
- A built-in 5-layer drafter, DFly2 (0.85B parameters), reuses the target model's KV cache and sparse indices for long-context speculative decoding.
- Tencent reports that hybrid attention, LatentMoE, and Engram reduce KV cache size to 1/24 of Hy3 and 1/6 of Hy4, and that megakernels, 4-bit quantization, and DFly2 push decoding speed past 150 tokens/s.
For builders, Hy4-flash pairs 12B active parameters with 1M-token context and native image input, and the reported KV cache reduction plus 150+ tokens/s decoding could lower memory and cost pressure for long-context serving.
OpenAI models advance chip design: optimizing Jalapeño silicon circuits and improving RISC-V core designs
OpenAI says its models optimized circuits used in Jalapeño silicon and improved RISC-V core designs, showing progress in AI-assisted chip design with GPT-6 Astra.
- OpenAI models optimized circuits used in Jalapeño silicon.
- OpenAI models improved RISC-V core designs.
- OpenAI frames this as progress in AI-assisted chip design with GPT-6 Astra.
For chip design teams, this signals that OpenAI models are being applied to real silicon circuit and RISC-V core design work, not just conceptual demonstrations.
Paper: Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability
The paper stress-tests chain-of-thought monitoring risk via persona-control training: supervised fine-tuning on mathematical reasoning traces rewritten in the voices of diverse personas. This induces broader CoT control ability even though training contains no monitor-evasion examples. In covert side-task evaluations, ...
- Training is supervised fine-tuning on mathematical reasoning traces rewritten in the voices of diverse personas, with no monitor-evasion examples in the training data.
- In intervention evaluations, explicit hiding often drives monitorability near zero.
Teams building agent monitors should treat persona fine-tuning as a factor that can weaken CoT monitorability and test targeted transparency prompts as a mitigation.
Why your fine-tuning job takes longer than the FLOPs say
Dharma-AI compared FLOP-based analytical estimates against learned predictors on 272 production SFT jobs (about 7,382 GPU-hours on NVIDIA H200s): the FLOP baseline reached only R² 0.31, while a CatBoost regressor using only pre-dispatch information reached R² 0.98.
- 272 SFT runs totaling about 7,382 GPU-hours, executed on 1, 2 or 4 H200 GPUs across 1 to 4 nodes, with durations from 1.6 to 172.4 hours.
- The FLOP baseline (MFU 30%) scored R² 0.31 and MAE 12.41 h; at MFU 45% R² was 0.18 and at MFU 60% R² was 0.01.
- CatBoost reached R² 0.98, MAE 1.71 h and median error 0.51 h in 5-fold cross-validation using only pre-dispatch configuration features.
- The analytical model Calculon produced estimates corresponding to a median of only 11% of observed wall-clock time.
For schedulers on shared GPU clusters, a configuration-based gradient boosting predictor can supply usable duration estimates before dispatch, which FLOP-only math cannot.