Bilingual

Qwen 3.6 27B is the sweet spot for local development
a month ago
- Qwen 3.6 27B is a powerful dense local model praised for punching above its weight, making it the first local model that works well as a general intelligence.
- The author recommends Qwen 3.6 27B over the faster mixture-of-experts 35B A3B variant for higher quality output, despite being slower, and demonstrates its capabilities in creative writing and coding tasks like creating a hexagonal minesweeper.
- Running Qwen 3.6 locally is straightforward with llama.cpp, supporting quantizations like 8-bit for reduced size and multi-token prediction for speed, and it can be integrated into tools like OpenCode for vibe coding.
- Performance tests on an M5 Macbook show Qwen 3.6 27B achieving around 30 tokens per second efficiently using GPU resources, with quantizations allowing it to run on devices with as little as 32 GB RAM.
- Benchmarks indicate Qwen 3.6 27B outperforms alternatives like Gemma 4 31B, and the author highlights the benefits of local models for privacy, fine-tuning, and avoiding reliance on subsidized proprietary APIs.
Model Training as Code
a month ago
- Model training complexity requires specialized teams and manual coordination is inefficient.
- Aleph Alpha's Savanna implements Model Training as Code (MTaC) for hermetic, one-click training runs.
- Manual training leads to human errors, lost learnings, and fragmented team ownership.
- MTaC offers composability, consensus via version control, and provenance through code history.
- Savanna uses trunk-based development, CI for validation, and immutable artefact versioning.
- It enables hyperparameter sweeps with caching to avoid redundant computations.
- The system runs on Flyte in Kubernetes, with artefact tracking and real-time monitoring.
- MTaC improves iteration speed, facilitates team restructuring, and supports auto-research.
AI is 'not smart' so what's next in artificial intelligence?
23 days ago
- Yann LeCun critiques current LLMs (like ChatGPT) for lacking real-world understanding and physical reasoning, comparing them unfavorably to even animals like rats.
- He left Meta to found AMI Labs, aiming to develop AI beyond LLMs, focusing on flexible systems for complex real-world tasks such as robotics and household chores.
- AMI Labs raised over $1bn in seed funding, with backers including Nvidia and Jeff Bezos's fund, highlighting investor confidence in next-generation AI.
- LeCun's approach, Joint Embedding Predictive Architecture (JEPA), creates abstractions to filter useless information and predict outcomes without over-predicting, like not guessing a pen's fall direction.
- Other researchers, like Ingmar Posner at Oxford, are working on 'World Models' that simulate environments to enable causal reasoning and explainable AI, inspired by advances since 2018.
- Projects like Google's Dreamer, DeepMind's Genie, and Wayve's Gaia exemplify the shift towards AI that learns from mental simulations, with applications in gaming and robotics.
- LeCun predicts AMI Labs' model will be tested in industrial settings soon, with long-term goals for general intelligence requiring minimal training.
- Despite AI advancements, humans will remain crucial for asking questions, creativity, and leadership, with AI acting as smarter assistants.
Show HN: Phobos – A tiny scale-free kernel language with tile-DAG support
19 days ago
- The author developed Phobos, a tiny kernel language inspired by Triton that compiles to PTX and runs on NVIDIA GPUs.
- Phobos achieves about 76% of cuBLAS SGEMM performance on an RTX 2080 SUPER, with optimizations like tiling and tensor core usage.
- The language abstracts thread-level details, focusing on tile-level operations for AI computations, enabling natural scaling to distributed clusters.
- The project involved learning low-level GPU programming, using MLIR and LLVM with Rust, and implementing autotuning for kernel parameters.
- A cluster prototype was built for distributed tile-DAG execution, though multi-node benchmarks were not performed due to hardware limitations.
Agent Harness Engineering
14 days ago
- Agent harness engineering focuses on the scaffolding built around AI models to create effective coding agents.
- A harness includes system prompts, tools, infrastructure, orchestration logic, hooks, and observability components.
- Harness engineering treats agent mistakes as signals to improve the system, adding constraints to prevent recurrence.
- Key harness components include filesystem and Git for state, bash and code execution as general-purpose tools, and sandboxes for safe execution.
- Memory and search mechanisms enable continual learning, while context management techniques combat context rot.
- Long-horizon execution is supported through loops like Ralph Loops, planning, verification, and planner/generator/evaluator splits.
- Hooks enforce rules and provide feedback, while AGENTS.md serves as a concise rulebook injected into prompts.
- Harnesses evolve with models; as models improve, harness components shift to address new capabilities and failure modes.
- The model-harness training loop shows that models become co-trained with specific harnesses, affecting performance.
- Harness-as-a-Service frameworks simplify agent development by providing pre-built runtime components for customization.
- Industry trends show convergence in harness patterns across top coding agents, indicating shared best practices.
- Future directions include multi-agent orchestration, self-improving harnesses, and dynamic tool assembly for tasks.
LambdaMART in Depth (2022)
12 days ago
- LambdaMART is a flexible ranking algorithm that allows optimization of different relevance metrics like DCG.
- It uses pairwise swapping to compute DCG impact (deltas) and learns from model errors via gradient boosting.
- The implementation in Python with Pandas involves computing deltas via self-joins, weighing by rho, and training decision trees iteratively.
- Key steps include feature preparation, delta calculation, error weighting with rho, lambda accumulation, and ensemble training.
- Learning rate and model size are important considerations to prevent overfitting and manage performance.
Detecting LLM-Generated Texts with “Classical” Machine Learning
9 days ago
- Mainstream LLM-generated text as of early 2026 shows strong statistical patterns, detectable by traditional machine learning models.
- An online demo using a simple SVM classifier achieves ~85% single-sentence accuracy, trained on human and LLM-generated data.
- Data was gathered from pre-2022 human texts and LLM-regenerated summaries via multiple low-cost/free APIs from models like Gemini, Qwen, and GLM.
- Seven binary classifiers were trained, using majority voting; detection remains robust against common anti-detection methods like translation or rewriting prompts.
- Testing on unseen models and human texts shows low false positives; analysis of Lofter content suggests significant undisclosed AI-generated material.
- The author expresses skepticism about LLMs' creativity, viewing AI-generated content as repetitive and shallow, and reflects on broader implications of AI misuse.
Mathematics of Data Science
9 days ago
- The book covers the mathematical foundations of data science, including topics like high-dimensional spaces, singular value decomposition, linear regression, graphs and networks, and deep learning.
- It addresses key techniques such as dimensionality reduction (both linear and nonlinear), optimization, classification, compressive sensing, and matrix recovery.
- The text also explores theoretical aspects like concentration of measure, Gaussian analysis, and matrix concentration inequalities, providing a comprehensive introduction to the mathematics behind data science.
How Does Pangram Work?
8 days ago
- Pangram is a classifier model that distinguishes AI-generated text from human writing by analyzing writing style patterns.
- It operates by mapping writing styles, placing similar authors close together; AI and human texts form distinct clusters on this map.
- Pangram evaluates entire texts, not just individual words, avoiding pitfalls of perplexity-based detectors which often misclassify historical documents.
- Training involves pairing human-written texts with AI-generated doubles to refine detection, focusing on subtle stylistic boundaries.
- The model requires a minimum word count for accuracy, and its performance is statistically validated through internal benchmarks and third-party testing.
Is AI Progress Real? Four Independent Metrics Show It
6 days ago
- AI benchmark improvements often face issues like contamination, saturation, and reward hacking, making it hard to distinguish real progress from test-specific gains.
- Four independent measures show a sharp acceleration in AI capabilities starting in late 2025, coinciding with releases like Grok 4.1, Gemini 3, Claude Opus 4.5, and GPT-5.2.
- METR's time horizon metric grew from 4 minutes to 12 hours in under two years, indicating rapid advancement in task execution for frontier models.
- TrackingAI's offline cognitive tests show AI systems went from outperforming fewer than 10% of humans to surpassing all but 1 in 44 within about two years.
- Humanity's Last Exam scores jumped from 2.7% to 53% in a year, reflecting exponential compute increases needed for Ph.D.-level knowledge.
- ARC-AGI-2 scores for abstract reasoning stalled for 15 months before surging in late 2025, with models like GPT-5.2 reaching 53% in weeks.
- Four compounding mechanisms drive improvement: pretraining efficiency gains, reinforcement learning with verifiable rewards, harness engineering, and a flywheel effect from AI self-improvement.
- Challenges like data exhaustion, reliability gaps in performance, and new benchmarks like ARC-AGI-3 (with interactive reasoning) could slow future progress.
- The multiproxy test—similar to climate science's hockey stick—corroborates AI acceleration across diverse, independent metrics, suggesting real, accelerating improvement.
Why non-invasive glucose monitoring is hard
4 days ago
- Non-invasive glucose monitoring is more difficult than blood oxygen monitoring due to glucose's low concentration (1000x less than hemoglobin), lack of distinct absorption peaks, and inability to use the AC/DC separation method.
- Various techniques have been explored: near-infrared (NIR) achieves ~20-25% MARD due to weak signals; mid-infrared (MIR) shows ~12% MARD but requires miniaturization; Raman spectroscopy reaches ~11.7-14.3% MARD but needs high laser power; photoacoustic spectroscopy is infeasible due to skin variability; fluorescence sensors like Eversense achieve ~8.5% MARD but are implantable; PPG with deep learning initially reports high accuracy but fails in large cohorts.
- Recent progress includes Apple's foundation model trained on 141,000 PPG/ECG waveforms, which successfully detected hypertension but remains unproven for glucose.
- The most likely path forward combines advanced machine learning with hardware advances (MIR or Raman), as indicated by Apple shrinking a tabletop prototype to iPhone size.
Intelligence from Learnable Novelty
4 days ago
- Intelligence appears in different fields as data compression, universal computation, or adaptive behavior, each with its own objective.
- Two influential drives—novelty search (seeking surprise) and the free-energy principle (avoiding surprise)—fail in mirror image ways, such as being transfixed by a noisy television or content in a dark room.
- Both failures stem from treating the surprise a learner can convert into knowledge and the surprise it cannot as the same quantity.
- The learnable part of information, called learnable novelty, yields the disparate projections of intelligence, and a closed-form estimator is built on a cheap, differentiable reservoir computer.
- As a measure with no supervision, the estimator recovers decades of complexity classification, ranking Turing-complete rule 110 highest among elementary cellular automata.
- As an objective, its gradient guides a neural cellular automaton from simple dynamics into soliton regimes, organizing image encoder representations around MNIST digit classes fully unsupervised.
- As an intrinsic reward for reinforcement learning, it supplies exploration missing from task rewards, improving on baselines in nine of ten environments without collapsing.
- Complexity generation, abstraction, and exploration emerge from ascent on one differentiable quantity, unifying the projections of intelligence on a common quantitative footing.
The linear algebra and calculus behind every model
2 days ago
- Every machine learning model relies on three pillars: linear algebra, calculus, and convex optimization.
- The MNIST dataset is treated as a matrix of 70,000 rows (images) and 784 columns (pixels) after flattening.
- Data exploration shows balanced classes, no missing values, and a bimodal pixel intensity distribution.
- Vectors have L1 and L2 norms; dot products measure similarity between images.
- Matrix multiplication forms a Gram matrix of pairwise similarities.
- Broadcasting allows adding a scalar bias to tensors without explicit loops.
- Projections decompose images into components along and perpendicular to the mean image direction.
- The Singular Value Decomposition (SVD) factors the centered data matrix and reveals 626 significant singular values out of 784.
The Open-Source Release of ML Video Codec (MLVC)
12 hours ago
- Microsoft announces open-sourcing of ML Video Codec (MLVC) under MIT License.
- MLVC is a neural-network-based video codec that uses less bandwidth than conventional codecs, improving streaming and call quality.
- It is the product iteration of DCVC family, with improved compression efficiency, real-time NPU performance, and cross-platform support.
- Compared to H.264, MLVC reduces bitrate by up to 87.8% at 360p and 82.7% at 540p; versus H.265, reductions are 75.5% and 65.4% respectively.
- MLVC is already being rolled out in Microsoft Teams for real peer-to-peer video calls.
- The release includes model source code, trained weights, training scripts, and platform conversion scripts.
- Contributions are welcome across streaming, communication, cloud gaming, AR/VR, and hardware domains.
- Future plans include stabilizing 540p real-time, expanding to 1080p, and building an open video codec ecosystem.
Becoming a Research Engineer at a Big LLM Lab
6 hours ago
- The author secured a research engineer role at Mistral after 18 months of strategic career development, emphasizing the difference between tactical (low-effort, high-return) and strategic (high-effort, compounding) actions.
- Personal timeline: defined goals, networked, upskilled (LeetCode, CS fundamentals), failed first application, quit job, joined Recurse Center, learned Rust, contributed to open source, wrote a paper, then applied in batches and got offers.
- Goal setting is crucial: define clear criteria for the ideal role (e.g., career inflection, rare skills, ownership, senior peers, IC path) to focus efforts and say no to compromises.
- Application playbook: batch applications, use network for referrals and insider info, schedule lower-stakes interviews first to gain practice.
- Networks rely on weak ties: build genuine friendships by doing interesting things, sharing them, and being open; online (Twitter) and offline (cohort programs) are effective.
- Career momentum requires building rare and valuable skills (career capital) and demonstrating an upward trajectory through projects, OSS contributions, publications, and CV entries.
- Interview types include initial screen, takehome, online assessment, coding interviews (LeetCode, API/system design, real-world), system design, culture fit, quiz, hiring manager, and reference checks—each with specific preparation strategies.
- The mental game involves managing rejection, uncertainty, and stigma; staying motivated by remembering the leverage and possibilities of job seeking.
- Tactics boil down to gaining an information advantage and acting on it, while strategy focuses on long-term skill building and proof points.

machine learning

Qwen 3.6 27B is the sweet spot for local development

Model Training as Code

AI is 'not smart' so what's next in artificial intelligence?

Show HN: Phobos – A tiny scale-free kernel language with tile-DAG support

Agent Harness Engineering

LambdaMART in Depth (2022)

Detecting LLM-Generated Texts with “Classical” Machine Learning

Mathematics of Data Science

How Does Pangram Work?

Is AI Progress Real? Four Independent Metrics Show It

Why non-invasive glucose monitoring is hard

Intelligence from Learnable Novelty

The linear algebra and calculus behind every model

The Open-Source Release of ML Video Codec (MLVC)

Becoming a Research Engineer at a Big LLM Lab