[One Million AI Engineer Special] How Stanford Teaches AI-Native Engineering — Mihail Eric, Monaco
What do you assign when a coding agent can finish last year’s homework in one prompt? In the second run of Stanford’s CS146S, Mihail Eric
The latest in AI, clustered and ranked. Repeated hype gets pushed down so the actual signal stays up top.
What do you assign when a coding agent can finish last year’s homework in one prompt? In the second run of Stanford’s CS146S, Mihail Eric
**Researchers from Alibaba and several Asian universities built a coding-agent benchmark from real user change requests. Even the best model fully completes only about one
_Goldman Sachs has restricted Anthropic's Claude Code and Microsoft's GitHub to its engineers, cutting off traders and bankers who used the tools to track client
| Metric | This model | Original model ([Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B](https://huggingface.co/Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B)) | | --- | --- | --- | | **Refusals** |
I don’t have a great segue into how it all happened and what I did in the early days, so here’s one of my larger
Discussion on evolving CS curriculum to include AI-native engineering and open-source contribution.
Introduction of Co-Design Bench, a benchmark showing coding agents struggle with real-world tasks compared to synthetic tests.
Report that Goldman Sachs has restricted employee access to Claude Code due to security concerns.
A tool for running large language models on consumer hardware with limited VRAM using offloading techniques.
Perspectives on the diminishing returns of scaling laws and the need for new learning paradigms.
Community discussion on practical local LLM workflows and performance expectations for Apple Silicon hardware.
A proposal for standardizing agent memory formats to improve observability and interoperability.
An argument regarding the current inability of LLMs to effectively architect complex AI systems.
Voters across the US turned against AI data centers on November 3, while AI use keeps climbing. Here's who gains when computing power gets harder to build, why the backlash could leave Big Tech strong
Overview of a project using LLMs to curate and filter historical Hacker News content into video format.
Discussion on the limitations of the computational theory of mind regarding AI consciousness.
Overview of world models for simulating physics in robotics and autonomous navigation.
Personal account of switching from Claude to Pi for daily AI tasks.
Music rights holders for Elvis Presley and Eminem have filed lawsuits against Suno over copyright infringement.
This video Opines about Arete-27B, post-trained over Qwen3.8-27B. 📬Weekly AI Newsletter: https://fahdmirza.substack.com/ 🔥 Buy Me a Coffee to support the channel: https://ko-fi.com/fahdmirza #arete2
This video locally installs and tests MiniCPM-OCR 1.0 (1.3B). It focuses on text-centric document understanding. 📬Weekly AI Newsletter: https://fahdmirza.substack.com/ 🔥 Buy Me a Coffee to support t
Study finds AI auditors effectively detect agent misbehavior but suffer from high false alarm rates.
Fine-tuning can effectively remove pro-state bias from large language models while preserving general capabilities.
Evaluation of nine AI agents on national lab research tasks shows high failure rates and frequent hallucinations.
Community fine-tune of Qwen 3.5, uncensored variant.
Grab details a layered defense strategy to manage stability risks from increased AI-generated code.
Viggle-Remake release for video-to-video character replacement and editing.
Gaussian splatting model for image-to-3D generation.
Moonshot AI Kimi K3.1 Pro model release for text generation.
Lightricks LTX-2.5 model for image-to-video and video-to-video generation.
Qwen 3.8 27B parameter model release for multimodal text generation.
NVIDIA Nemotron 3.5 30B model optimized for speech and transcription tasks.
LiquidAI releases LFM3-1.7B, a small language model optimized for edge deployment.
A reflection on the generalization of the Bitter Lesson in the context of modern AI development.
Commentary on the limitations of relying on AI-generated code without understanding the underlying implementation.
Analysis showing Claude Code's automation features led to accidental credential exposure in public GitHub repositories.
Technical analysis of methods for detecting and mitigating prompt injection attacks embedded within PDF documents.
Fine-tuned Qwen 3.5 4B model specialized for Three.js and 3D coding tasks.
NousResearch releases Arete-27B, a fine-tuned model focused on value alignment.
Black Forest Labs releases Flux-3 image generation model optimized for ComfyUI.
Unsloth provides GGUF quantization for the Qwen3.8-Flash-Next model.
Preview release of a 65B parameter model with custom architecture.
Fine-tune of the MiniMax-H3 base model.
AtomicChat releases an uncensored GGUF quantization of DeepSeek-V4.1-Flash.
Numbat model release featuring 8-bit quantization and custom code implementation.
TwIL-8B model release based on Granite architecture, provided in GGUF format.
PrunaAI releases P-H3-Image-Comfy, a diffusion model variant for image generation workflows.
Benchmark data suggests the performance gap in cyber-attack capabilities between open and closed models is narrowing.
Reports suggest inadequate vetting of Chinese firms receiving Nvidia H200 chips.
This video tests Inkling 2 Small, a MoE with 594B total parameters. 📬Weekly AI Newsletter: https://fahdmirza.substack.com/ 🔥 Buy Me a Coffee to support the channel: https://ko-fi.com/fahdmirza #inkl
Anthropic reportedly plans an IPO at a valuation between $1.25 and $1.75 trillion.
Report on Anthropic's internal use of AI to accelerate model development cycles.
Podcast discussion on the current state of cybersecurity in the age of AI.
Same signal-first ranking, earlier dates.
Open archive