Welcome to the world of AGI.

WER benchmarks claiming 4% for streaming and 2.6% for non-streaming use-cases are impressive. Curious if they hold up in real world scenarios. Additionally, would love more details on speaker diarization techniques. It's been the hardest thing to get right for longer multiple-speaker speech segments.

This is great for Linux and the open source community! People like @dhh are forces for good and so much needed in today’s world.

Go 1.27 was released today, bringing a couple of neat updates to the language: 1) generic methods -- a method can now declare its own type parameters. So, instead of: func (r *Rand) Int32N(n int32) int32 func (r *Rand) Int64N(n int64) int64 func (r *Rand) IntN(n int) int you can replace them with just: func (r *Rand) N[Int intType](n Int) Int 2) struct literal keys can now be any valid fields from the embedded structs 3) function type inference -- generic functions can now be used without explicit type arguments in composite literals, type conversions, and channel sends 4) uuid package -- provides native support for generating and parsing UUIDs 5) experimental SIMD support in a new simd package

AI Labs: AI will cure cancer. Moderna: hold my beer.

AI infra (both physical and software) will rule the next 5+ years.

The Qwen3.8 family of models is the closest to the frontier for agentic capabilities. The most exciting part is that you can download the weights, self-host the model and not lose much in terms of capabilities. This means: - full ownership of your data - flexibility to fine-tune the model - control over the lifecycle from generation to application to improvement

Make something agents want.

The companies are realizing that the AI tools themselves are not enough. Necessary, but not sufficient. That’s why we’re seeing an explosion of investment in AI/ML platforms, agentic architectures and obviously eval frameworks and high quality data curation. What a time to be an engineer!

Over the last 15 years we've witnessed unparalleled innovation in AI: 1/ Model Architectures: AlexNet (2012), Word2Vec (2013), Seq2Seq (2014), Transformers (2017), Visual Transformers (ViT) (2021), MoEs (2021) 2/ Training algos: unsupervised and self-supervised learning, asynchronous training, distillation, SFT, RLHF 3/ Inference algos: Chain-of-Thought, PagedAttention, Continuous Batching, Speculative Decoding, Quantization 4/ Software: scikit-learn, XGBoost, TensorFlow, PyTorch, Jax, vLLM, SGLang, NVIDIA Triton 5/ Hardware: TPUv1 (2013) -> Ironwood (2025), NVIDIA Tesla K80 (Kepler) (2014) -> B200 (Blackwell) (2025)

There’s been a lot of discussion about the proper agent sandboxing and how to avoid agents “escaping” the sandbox and hacking other sites, or worse share your private keys and secrets outside. My take is that the container is not the sandbox. The environment is. A thousand agents with perfect per-agent isolation still share the package registry, DNS, and the secrets vault, and that shared layer is where escapes actually happen. Instead, focus on building a secure cloud environment (a separate Google project or AWS account dedicated to agents) with strong network isolation (secure, firewalled VPC perimeter), a private package registry and a single egress gateway which only allows access to outside inference endpoints. Once you solve the isolation at the environment level you can safely run agents on any compute inside it: VMs, containers, serverless, etc.

I'll just leave it as is. What a year...

PSA: California recently launched DROP, a free tool that lets you delete your personal data from hundreds of data brokers with a single request. They started processing requests on Aug 1 and data brokers have 90 days to completely delete your data. Submit your DROP request today if you haven't already. privacy.ca.gov/drop

It’s fascinating to see how far open-weight models have come since the original DeepSeek release in early 2025. The latest Qwen-3.8 is not just crushing the benchmarks, but it’s also significantly cheaper than Fable 5.

The most recent versions of Firefox support something called "Tab Containers" which let you separate your tabs into different isolated environments with their own cookies, history and data access, preventing cross tab-communication between different containers. Imagine having two Google accounts: one for work, one personal. Instead of starting two different browser sessions, now you can just group the websites in different containers and manage them from within your existing session accessing them as separate tabs. Pretty neat feature @firefox team!

LinkedIn just shipped a report button for this. 𝕏 seems to have quietly handled it in ranking instead. Or I fixed my own following list, hard to tell from the inside.