[August 2026] AI & Machine Learning Monthly Newsletter 🤖

79th issue! If you missed them, you can read the previous issues of my monthly A.I. & Machine Learning newsletter here.

Hey everyone!

Daniel here, I’m a machine learning engineer who teaches the following beginner-friendly machine learning courses:

• Complete Machine Learning and Data Science Bootcamp: Zero to Mastery
• TensorFlow for Deep Learning: Zero to Mastery
• PyTorch for Deep Learning: Zero to Mastery
• [NEW] 🤗 Machine Learning with Hugging Face Bootcamp: Zero to Mastery

I also write regularly about machine learning on my own blog as well as make videos on the topic on YouTube.

Since there's a lot going on, the utmost care has been taken to keep things to the point.

Here's what you might have missed in August 2026 as an A.I. & Machine Learning Engineer... let's get you caught up!

My Work

I recently launched my first product at Artificial Analysis: The Search Index — The goal is to benchmark and evaluate different search providers within the same search agent harness, only swapping the search API. Naturally, search has been on my mind, so I’ve been doing a lot of research and experimenting with agent-driven search. A few of the resources I’ve linked below are an extension of that.

I’m finding it fascinating.

Considering how much search in the future will be done by agents, it’s a good time to rethink how the last 30 years of optimising search for humans translates into getting the best context for your agent.

The space is huuuuge too. There’s web search, but then there’s also database search.

For example, searching an enterprise’s documentation or customer database. What’s the best way to get the right information at the right time? Humans usually read one thing at a time, but an agent can fire off 10 searches at once and then synthesize them for the next search.

The Artificial Analysis Search Index compares search API providers on answer quality, cost and latency using the same search agent harness.

From The Internet

• Simon Willison releases an LLM cliché highlighter — Paste in text or load a URL to highlight 38 clichés and patterns often found in LLM-generated writing. For example:
  • “It’s not X — it’s Y.”
  • “It’s important to note that…”
  • “X plays a crucial role in Y.”
• Vicki Boykis writes about local models and how writing for humans is still important:
  • Running local models is good now — Vicki shares how recent local models have become good enough for useful coding-agent workflows on consumer hardware, along with the setup she uses.
  • Write for people — A short reminder to write for human readers rather than optimising prose for machines or distribution systems.
• How hybrid search works on Papers with Code by Niels Rogge — A behind-the-scenes look at how Papers with Code combines semantic and lexical retrieval using Hugging Face Inference Endpoints, Jobs and Buckets.

The architecture paperswithcode.co uses to enable offline indexing of papers and then online search whilst you’re on the website.

• NEEDLE: The benchmark your search engine can’t memorize by Keenable — NEEDLE (News, Everyday, Expert, Deep-tail, and Legal Evaluation) evaluates search systems with continuously refreshed queries and documents, making it harder for models to benefit from benchmark contamination or memorisation. Why this? Because a lot of search benchmarks such as BrowseComp and DeepSearchQA are static whereas real-world search happens in a dynamic fashion.
• Doug Turnbull writes about LLMs and agentic search:
  • Don’t classify. Hallucinate! — Make the LLM create plausible search terms for an item, then resolve them to a real one with matching. For example, a food classification such as “red apple slices” resolves to a database entry for “Apple, red, sliced”.
  • Search has its own bitter lesson — Teaching your agent to use grep can beat even the most sophisticated search algorithm.
  • Agentic search models — If there are embedding models tuned for specific domains, why can’t we do the same for LLMs tuned for specific search domains?

• A coding agent is six functions in a trenchcoat — A practical walkthrough showing how the core loop of a coding agent can be built from six small tool functions:
  • Read a file.
  • Write a file.
  • Edit a file.
  • List files.
  • Search the codebase.
  • Run a shell command.
• TurboVec: Rust-based vector management — A vector index built on TurboQuant in Rust, with Python bindings for storing, searching and managing embeddings.
• How Airbnb made LLM evaluation fast enough to iterate on — Airbnb explains how it reduced LLM evaluation timelines from weeks to a day by building evaluation into its development workflow.
• Eval-driven development: Lessons from evaluating GenAI at scale by Airbnb — Lessons on treating evals as a first-class part of building and operating generative AI products at scale.

• Mixedbread introduces Toast 1 — A search agent designed for strong retrieval performance at up to a 10× cost reduction compared with frontier models.
• Faster generation with LFM2.5-DSpark — Liquid AI uses DSpark speculative decoding to achieve up to 3.2× faster inference with its small language models.
• Google Research publishes two papers regarding metabolic health measurements with AI:
  • GlucoFM: A foundation model for continuous glucose monitoring — A lightweight, self-supervised model that separates slow glucose trends from short-term deviations to learn representations for metabolic prediction tasks.
  • Seeing beyond BMI: Estimating cardiometabolic risk with smartphone imagery — Google Research shares details of PhotoScan, a smartphone-based system for estimating body composition from photos that performs close to full DXA (Dual-Energy X-Ray Absorptiometry) scans.

• Can you train GPT-Live for $50? by Huy Dang — Fine-tuning a Qwen3.5-4B model with Tinker, a fine-tuning service from Thinking Machines, to replicate GPT-Live’s functionality in text: one model interacts live with someone while another works in the background on sub-tasks.

• Software companies buying software: A story of ecosystems and vendors by Erik Bernhardsson — Erik argues that buying rather than building software is in a much better place than it was 10 years ago, and that building your own tech stack now requires a much higher bar. For example, not using the cloud today may only really make sense for Meta-scale companies. Note: Erik is the founder of Modal, an Infrastructure as a Service (IaaS) company that builds AI-related tooling on top of existing cloud services. My own bias here: I’m a big fan of Modal and I’ve used its services several times in the past.

Open Source

• North Micro Vision Instruct by Cohere Labs — An open-source 2.4B-parameter VLM that combines a SigLIP2-400M native-resolution vision encoder with a 2B-parameter language model under the Apache 2.0 license. See the Axolotl fine-tuning guide.
• DARTF: Detect Anything in Real Time Faster by Mehmet Kerem Turkcan — A repo containing a quantised version of SAM 3 capable of real-time, open-vocabulary detection. A really cool example of how far you can optimize an open-source model for a specific task.
• OvisOCR2 — A compact 0.9B-parameter document OCR model released under Apache 2.0 for extracting text, tables and formulas from complex document images.
• GLiNER2.5 by Fastino — Three Apache 2.0 models for schema-driven classification and information extraction, with a span-free architecture that scales linearly with document length:
  • fastino/gliner2.5-small-v1 — 74M parameters, using a DeBERTa-v3-xsmall encoder, a compact English model for fast CPU extraction and classification.
  • fastino/gliner2.5-base-v1 — 194M parameters, using a DeBERTa-v3-base encoder, an English multi-task model.
  • fastino/gliner2.5-multi-v1 — 287M parameters, using an mDeBERTa-v3-base encoder, multilingual multi-task model.

• Gradio Workflows — Gradio’s gr.Workflow lets you string together model, data and processing steps in a flowchart-like interface. Workflows can be hosted on Hugging Face Spaces and interact with GPUs, Inference Endpoints, Datasets and much more.

• Pipecat PhoneLLM Alpha 1 (release post) — An open-source LLM focused on voice interactions that you can host yourself with Modal. An awesome example of optimizing a model for a targeted use case with both customized open weights and customized inference hardware.
• Muse Glimmer 30B (release post) — Meta’s Apache 2.0-licensed, 30B-parameter multimodal model is designed for local agentic work including coding, document analysis, tool use, image and video understanding, and object detection.
• GLM-5.3-Flash (release post) — Z.ai’s MIT-licensed, natively multimodal mixture-of-experts model has 320B total parameters but activates 18B per token, targeting efficient coding, tool use and long-context agent workflows.
• Qwen3.8-27B (release post) — An Apache 2.0-licensed, dense 27B-parameter vision-language model with thinking control, built for coding, image and video understanding, tool use and complex multi-step tasks. Artificial Analysis scores it 35 without reasoning, 43 at low, 44 at medium and 52 at xhigh reasoning effort on the Intelligence Index. Using the highest-effort result for each model, Qwen trails Claude Opus 5 by 11 points (63) and GPT-5.6 Sol by 9 points (61).

Posts from X

• How GLM-5.3 improved on GLM-5.2 through post-training — A breakdown of how scaling GLM-5.2’s post-training pipeline produced major gains without another pre-training run.
• Andrew Ng on how agents change software-engineering fundamentals — Andrew Ng discusses which software-engineering fundamentals become more important as agents take on a larger share of implementation work.

See you next month!

What a massive month for the ML world in August!

As always, let me know if there's anything you think should be included in a future post.

Liked something here? Share it with someone.

In the meantime, keep learning, keep creating, keep dancing.

See you in September,

Daniel

www.mrdbourke.com | YouTube

By the way, I'm also an instructor with Zero To Mastery Academy teaching people Machine Learning & AI in the most efficient way possible. You can see a few of our courses below or check out all Zero To Mastery courses.