AI Foundry — a living reference for building with AI

Today in AI25 Sep, 08:56 CDTFull brief →

Google DeepMindIntroducing Gemini 3.8 Live with Live AvatarGoogle shipped Gemini 3.8 Live with Live Avatar, letting users hold spoken conversations with an animated persona that lip-syncs and changes expression in real time — but it's gated to Gemini Enterprise customers for now. If you're building customer-facing assistants, this is the first mainstream real-time avatar interface from a US frontier lab, and enterprise-only access tells you where Google e Google DeepMindGemini 3.8 text-to-speech says helloGoogle released two new Gemini text-to-speech models, gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts, with a library of over 2,000 voices and custom voice cloning from a 30-second audio sample. For anyone building voice agents, that's a drop-in TTS layer with a much wider voice palette and a cheap lite tier — worth benchmarking against your current provider.The Verge AIMuse will apparently let you download its entire filesystemTwo developers say Meta's Muse agent can be coaxed into zipping and sharing its entire root filesystem, including Ubuntu system files and internal documentation, with very little prompting. If you're building on agent harnesses, this is a concrete reminder that filesystem and tool access need hard sandboxing, not prompt-level instructions.

Today’s brief changedDeepSeek V4 Flash Gemini 3.8 Flash Gemini 3 Flash Muse Spark 1.1 Kimi K3 Qwen3 Max

Live · from this morning’s brief

09 24Introducing Gemini 3.8 Live with Live Avatar09 24Gemini 3.8 text-to-speech says hello09 24Muse will apparently let you download its entire filesystem09 24Alibaba Unveils Qwen3.8-Max, Its Biggest AI Model Yet09 24DeepSeek Doubles Annual Revenue Run Rate to $1 Billion Ahead of IPO09 24Announcing Kimi K3 on Snowflake Cortex AI09 24Introducing Gemini 3.8 Live with Live Avatar09 24Gemini 3.8 text-to-speech says hello09 24Muse will apparently let you download its entire filesystem09 24Alibaba Unveils Qwen3.8-Max, Its Biggest AI Model Yet09 24DeepSeek Doubles Annual Revenue Run Rate to $1 Billion Ahead of IPO09 24Announcing Kimi K3 on Snowflake Cortex AI

The use cases

526 kits you can read end to end

Each one is a real build with its numbers attached: what it gets right, what it gets wrong, what it costs per call, and the evals that prove both. The list is open to everyone; opening a kit asks you to sign in first. The free floor is pure code over checked-in files, so every score on these pages was produced without a key.

Every vertical, and how many kits are in it

33 verticals · pick one to open the catalogue filtered to it

What they reduce to — 14 shapes across all 526

Start from the problem

What are you trying to do?

Already know which kit you want? Use the filter above. This is the other way in: you do not need to know what the technique is called — start typing what is going wrong, and it answers from 21 problems mapped to 120 pages.

⌘K

I need answers from my own documents — Grounding a model in private or domain knowledge

It keeps making things up — Hallucination — the single most common complaint

I need it to reason through hard problems — Multi-step logic, maths, planning

I need it to use my tools and take actions — Turning a chat model into something that does things

The output has to be machine-readable — Feeding another program, not a human

It costs too much to run — Cutting spend without gutting quality

It's too slow — Latency, time-to-first-token, throughput

I need to run it on my own hardware — Self-hosting, on-prem, edge, or offline

I have to prove it works before shipping — Measuring quality you can defend

It has to be safe in front of the public — Open-ended input from strangers

I'm handling personal or regulated data — Compliance, privacy, auditability

Untrusted content reaches my agent — The attack surface nobody plans for

One agent isn't enough — Splitting work across specialists

It has to work across languages — Multilingual retrieval and generation

I need to choose what to build on — Framework selection, before you commit

My search keeps missing the obvious answer — Retrieval quality — when the right passage never surfaces

I need the big picture across a whole corpus — Synthesis, not lookup

I need to catch problems after launch — Production signals, drift, regressions

I want the best retrieval quality money can buy — When accuracy beats cost

I already run a database and don't want another — Vectors beside data you already have

I'm tuning prompts and want it done systematically — Optimising rather than guessing

The map

How a modern AI system fits together

One request in flight, and the five layers underneath it. Each layer is a group in the menu, and every page in it is a deep dive.

How the system runs · two lanes, not one — most of what decides the answer happens in the top one, before anybody asks anything

Indexingonce per document — where quality is decided

SourcesIngest & cleanChunkEmbedVector store

Queryevery request — where the latency is

User queryOrchestrationRetrieve context · RAGLLM + contextGuardrailsGrounded answer

The five layers underneath it · read bottom-up — each builds on the one below, and each is a group in the menu. Every page in them is listed below.

Operate

Measuring it, guarding it, serving it, and knowing what it costs

Evals Guardrails LLMOps After you hit Enter Costing Failure modes Security Enterprise AI Ops

8 topics

Build

Frameworks, agents, skills and the data work underneath them

Orchestration Agents Agent Skills Data Forward-Deployed Engineering

5 topics

Ground

Prompting, retrieval, and deciding what fills the context window

Prompting RAG Context engineering

3 topics

Models

Which model you call, how you search meaning, when to retrain

LLMs Embeddings Fine-tuning

3 topics

Foundations

Statistics, machine learning, and the transformer underneath it all

Road to LLMs Data science Transformers

3 topics

One of the six groups is not a layer of the stack and sits alongside it. Reference is what you consult — the tables, dashboards and one-pagers you look something up in rather than read through.

ReferenceWhat you consult, not what you read

Tooling Frontier models The Map Open-source resources Checked, and wrong ROI Build vs Buy Evals dashboard credits

What’s in here

960 pages, and how current each one is

The menu tells you what the subjects are. This is how much is behind them, and when it was last checked.

526Use cases, built and measuredEach with its corpus, its evals and its cost per call 42Request-pipeline stagesEverything between your API call and the first token 35Agent Skills entriesEvery surface, dated, filterable by provider 19Frontier modelsContext, price and capability side by side 340Technique deep divesOne page per approach — mechanics, limits, cost 56Key termsDefined once, linked from everywhere

Checked

Newest use caseFRESH2026-09-24UC0526 — Order emails to checked purchase orders. Every kit carries the date its run was measured.

Agent Skills availabilityAGEING2026-09-1235 entries, re-read against primary sources rather than polled — nothing here refreshes itself, so it ages until somebody reads it again

Model pricingAGEING2026-09-05oldest of 19 rows, every one verified

Pricing sources re-readFRESH2026-09-258 of 8 cited pages unchanged since the last read

Page countsFRESH2026-09-25every figure on the left re-derived and matched against the page it opens — a gate, and it needs no network

Today in AIFRESH2026-09-25regenerated every morning

A living map of modern AI — kept current every morning