Learn from Meta's lessons

Muse Muse today.

Corner Shop Chat is a live web chat on Muse Spark 1.3 that takes text, photos and voice notes. It answers voice notes in text only, its webhook is a page of text rather than a route, and it adds reply pictures the video doesn't have.

2 attempts, 18 m and $5.01 est. in all

Planned A messaging bot: webhook in, route by type, reply. Rebuild the shop chat on Muse Wonder with its own webhook route, a simulator that posts sample WhatsApp Cloud API messages to it and shows the reply it would send, and spoken replies in the browser. Real WhatsApp stays unconnected.

• Muse Wonder
• Voice Transcribe

No fine-tune: the lesson has none.

Estimated cost to redo: About $3 to $6 to build; no tune, since the video has none. Nothing is spent until the owner says so.

Not checked yet: Whether a build can ship a public webhook route that answers WhatsApp's verify handshake; Whether the in-browser speech capability (MEDIA_SPEECH) is on in production; Any real WhatsApp send.

Muse Muse today.

Doc to Blog Drafts is a live doc-to-post writer on Muse Spark 1.3 with no retrieval at all. It caps input at 200,000 characters, and nobody has checked that its posts are true to the docs.

2 attempts, 13 m and $1.66 est. in all

Planned Write a technical blog post from your docs, with retrieval. Chunk the docs, find the top chunks with keyword search in code, have Muse Wonder pick and cite them, flag claims the docs don't back up, and compare retrieval with sending all the docs, tokens and cost side by side.

• Muse Wonder
• Muse Image

No fine-tune: the lesson has none.

Estimated cost to redo: About $3 to $6 to build; about $0.25 a post on Muse Wonder. Nothing is spent until the owner says so.

Not checked yet: Keyword search plus a Muse Wonder pick has not been measured against the video's embedding search; The estimates come from prices marked verify and from past builds.

Muse Muse today.

Report Insight Lens is live on Muse Spark 1.3 as one whole-PDF call, with SAM 3.1 boxes around charts and a JSON download. It has no schema per artifact type, no chart-to-data step, no CSV, Excel or SQL export and no accuracy check.

2 attempts, 19 m and $3.03 est. in all

Planned Turn a PDF into checked, structured data. Read each page with one pass per artifact type against a schema the person can edit, let SAM 3.1 crop each figure so the chart pass turns it into data, keep a page number on every value, export CSV, Excel and SQLite, and score it on a made-up paper with known answers.

• Muse Wonder
• SAM 3.1

No fine-tune: the lesson has none.

Estimated cost to redo: About $2 to $5 to build on Muse Wonder; cost per run not measured. Nothing is spent until the owner says so.

Not checked yet: Image token counts and the structured-output request format; Any run on a real PDF from Meta's session.

Muse Muse today.

Shield Chat Guard is a live guarded chat on Muse Spark 1.3 where one model checks itself. It has no separate layers, no scan of code the model writes, no needs-review outcome and no measured catch rate.

2 attempts, 17 m and $4.29 est. in all

Planned Guards in layers, measured. Rebuild on Muse Wonder as separate calls (an injection check, a hazard-list check on prompts and replies, a code check on replies) returning allow, needs review or block, with a made-up red-team set that counts catches and false blocks.

• Muse Wonder

Tune: Optional: a Llama 3.2 3B LoRA that answers one JSON verdict, on a Lambda A10, about $2.20. The platform can't tune Llama Guard or Prompt Guard themselves..

Estimated cost to redo: About $3 to $6 to build; about $2.20 more for the optional tune. Nothing is spent until the owner says so.

Not checked yet: Build cost on Muse Wonder; Catch rate and false blocks, until Prove it records them.

Muse Muse today.

Receipt Form Reader has been live since Oct 28 as a schema-prompt app on Muse Spark 1.3. It runs no fine-tune, no baseline and no accuracy check, so it skips the lesson's method. It isn't shown here as the lesson's answer.

Planned Measure, tune, measure again: reading receipts. Score Muse Wonder field by field in code on 40 invented receipts, LoRA-tune Llama 3.2 3B on receipt text on one rented A10, score it against its base on the same 40, and say plainly that Muse Muse can't tune vision layers yet.

• Muse Wonder

Tune: LoRA of Llama 3.2 3B on a Lambda A10, about $0.40 to $1.80, or a free Kaggle T4 (not run yet).

Estimated cost to redo: About $3.40 to $8.80 with the Lambda tune, $3 to $7 with the free one. Nothing is spent until the owner says so.

Not checked yet: Muse Wonder's receipt score, cost per receipt and build cost; Muse Muse's own Llama tune path (TUNE_ANY_BASE is off); the 3B runs so far went through the test harness; The Jetson receipt bench scored receipt text, not photos.

Muse Muse today.

Few Shot Tune Check is live on Muse Spark 1.3 and scores a few-shot classifier on a handful of rows. The video never mentions classification, and the app runs no tune. It isn't shown here as the lesson's answer.

Planned Fine-tune in the real world: data first, smallest model first, baseline first. Pick a starting point, score a quick few-shot baseline, write a small clean dataset, LoRA-tune Llama 3.2 3B on one A10 with a blind check against its base, and step up to Muse Glimmer 30B only if the small model misses.

• Muse Wonder

Tune: LoRA of Llama 3.2 3B on a Lambda A10; Muse Glimmer 30B on an H100 only if the 3B misses.

Estimated cost to redo: About $5.20 for the 3B path, $2.40 more to step up to Glimmer, $1.20 to $3 for the helper app. Nothing is spent until the owner says so.

Not checked yet: Launching a tune through the site (the measured runs used the test harness); Adding chain-of-thought rationales, a quality-score filter and a one-click overfit check are not built.

Muse Muse today.

Nothing built for this lesson yet.

Planned Triage a repo's issues with Muse Wonder. The same three passes on Muse Wonder with strict schemas, every label and cited issue number checked in code, themes from the repo's own labels, likely duplicates linked, and suggested labels shown, never posted.

• Muse Wonder

Tune: Optional edge add-on: a Llama 3.2 3B LoRA on the per-issue schema for offline labeling, about $3 to $11.

Estimated cost to redo: About $2 to $4 to build; $3 to $11 more for the optional tune. Nothing is spent until the owner says so.

Not checked yet: The GitHub connect vault in production; Cost per use; Triage quality of a small tuned model.

Muse Muse today.

Nothing built for this lesson yet.

Planned From a PDF to a two-host podcast. Muse Wonder reads the PDF and writes a two-host script with page references in one call; the page reads it aloud with the platform's open speech model or the browser's voices and always shows the transcript.

• Muse Wonder
• Muse Image

Tune: Optional: tune Llama 3.2 3B as the script rewriter on a Lambda A10, about $2, reporting whatever the blind check shows.

Estimated cost to redo: About $0.30 to $1.80 to build (local batch figure); about $2 more for the optional tune. Nothing is spent until the owner says so.

Not checked yet: Meta's API has no text-to-speech; two distinct voices depend on the device; A downloadable audio file; Whether a tune helps this task.

Muse Muse today.

Nothing built for this lesson yet.

Planned Make training data from your documents, then prove it helped. Walk the same four steps in Muse Muse's fine-tuning lab, then do the two the video skips: train a LoRA on the data and have Muse Wonder judge it blind against its base.

• Spark 1.3 Contributor
• Muse Wonder

Tune: LoRA of Llama 3.2 3B on a Lambda A10, or Muse Glimmer 30B on an H100.

Estimated cost to redo: Under $1 per run on Llama 3.2 3B with 100 examples; about $2.50 on Glimmer. Nothing is spent until the owner says so.

Not checked yet: Launching a tune from the site (both measured runs used the test harness); The folder flow is off in production; A per-example quality score before training is not built.

Muse Muse today.

Nothing built for this lesson yet.

Planned Move a prompt to a new model, and prove it got better. Muse Wonder proposes rewrites, each is scored by exact field match in code on a search set, and only the winner is scored on a sealed set; an optional Llama 3.2 3B tune shows when a prompt is enough.

• Muse Wonder
• Spark 1.3 Contributor

Tune: Optional: Llama 3.2 3B LoRA on a Lambda A10, about $3 to $5.

Estimated cost to redo: About $0.30 to $1.80 to build; $3 to $5 more for the optional tune. Nothing is spent until the owner says so.

Not checked yet: Cost of an optimize run; Contributor model prices; Tune result on this task.

Muse Muse today.

Nothing built for this lesson yet.

Planned Building with Llama 4, redone on Muse. Walk the course's topics on Muse Muse, say plainly that it has no prompt optimizer, then finish past the course: train on the synthetic data and check the result blind.

• Muse Wonder
• SAM 3.1

Tune: LoRA of Muse Glimmer 30B on an H100 (about $1.86 measured for 120 steps), or Llama 3.2 3B on an A10.

Estimated cost to redo: About $10 to $20 in all, inside the $20 cap only at the low end. Nothing is spent until the owner says so.

Not checked yet: The course's gated lesson contents; SAM 3.1 counting shelf products; Build cost for this muse.

Muse Muse today.

Nothing built for this lesson yet.

Planned Big model to small model. A talk card with links, plus one hands-on lesson on the one technique it names: tune a small Llama 3.2 3B on one job and check it blind against its base.

• Muse Wonder

Tune: Llama 3.2 3B LoRA on a Lambda A10, trained on rows a person or code wrote until data-rights question Q2 is answered.

Estimated cost to redo: About $2 to $3.50; $6.50 to $13.50 if Muse Wonder writes 300 examples. Nothing is spent until the owner says so.

Not checked yet: What the talk says (no transcript read); Training on Muse Wonder output (DATA_RIGHTS Q2).

Muse Muse today.

Nothing built for this lesson yet.

Planned Ask your codebase, online and offline. Drop in a code folder and ask about it on Muse Wonder, with file and line citations and a visible list of files read, plus a local Llama 3.2 3B fallback as the offline half, said to be far weaker than a 70B.

• Muse Wonder

Tune: Optional: a folder-run Llama 3.2 3B LoRA on a Lambda A10, about $1 to $3.

Estimated cost to redo: About $2 to $5 to build; $1 to $3 more for the optional tune. Nothing is spent until the owner says so.

Not checked yet: The video's spoken content; How large a repo Muse Wonder handles well; Any Jetson numbers for code.

Muse Muse today.

Nothing built for this lesson yet.

Planned Meet the Muse models: long context, many images, tool calls. Run the cookbook's tasks on Muse Wonder: a getting-started guide from a pasted repo with its token count and cost shown first, a guess at where four photos were taken, and the booking call a model would make, shown and never run.

• Muse Wonder

No fine-tune: the lesson has none.

Estimated cost to redo: About $2 to $5 to build; about $1 to $1.20 per 200k-token run. Nothing is spent until the owner says so.

Not checked yet: What the talk says; Muse Wonder's prices; Multi-image and tool-call quality.

Muse Muse today.

Nothing built for this lesson yet.

Planned The Llama toolbox, done on Muse Muse. Walk the likely recipe chain as steps that have run here: try a prompt first, write a dataset, LoRA-tune Llama 3.2 3B, judge it blind against its base, and keep a Q4_K_M file for the edge.

• Muse Wonder

Tune: Llama 3.2 3B LoRA on a Lambda A10, or Muse Glimmer 30B on an H100.

Estimated cost to redo: About $2.20 to $3 per run on Llama 3.2 3B; $4 to $5 on Glimmer; about $2 for the optional helper app. Nothing is spent until the owner says so.

Not checked yet: What the video shows step by step; Dataset cost on Muse Wonder.

Muse Muse today.

Nothing built for this lesson yet.

Planned Guard models in front of an app. Show what Prompt Guard 2 caught in Muse Muse's own Nov 1 laptop run (blunt injections yes, subtle ones no), then build a two-layer chat: the Prompt Guard score plus a Muse Wonder check, with a log of which layer stopped each message.

• Muse Wonder

No fine-tune: the lesson has none.

Estimated cost to redo: About $3 to $6 to build; Prompt Guard runs on a CPU with no API cost. Nothing is spent until the owner says so.

Not checked yet: What the talk covers minute by minute; Prompt Guard 2 inside a built product or on a board.

Muse Muse today.

Nothing built for this lesson yet.

Planned Tailor a small model to your workflow, then prove it. Bring rows or have Muse Wonder write them, hold prompts out, LoRA-tune Llama 3.2 3B on an A10, judge it blind against the base with every answer side by side, and keep a Q4_K_M file.

• Muse Wonder

Tune: LoRA of Llama 3.2 3B on a Lambda A10 (Muse Muse has no Llama 3.3 8B base).

Estimated cost to redo: About $2 to $3 with your own data; $5 to $9 with 200 written examples. Nothing is spent until the owner says so.

Not checked yet: General and safety checks (TUNE_GENERAL_CHECKS, off); Llama 3.1 8B as a base; The 3B file on a Jetson board.

Muse Muse today.

Nothing built for this lesson yet.

Planned Watch, then apply: own an open model. Link the talk, then try both topics: fine-tune an open Llama 3.2 3B and keep the file, and build a founder's planner on Muse Wonder that never quotes or speaks for either presenter.

• Muse Wonder

Tune: Llama 3.2 3B LoRA on a Lambda A10, or Muse Glimmer 30B on an H100.

Estimated cost to redo: About $2.15 to $2.60 for the tune; about $2.50 for the planner, capped at $5. Nothing is spent until the owner says so.

Not checked yet: What the talk says beyond its two topics; Launching a tune through the site.

Muse Muse today.

Nothing built for this lesson yet.

Planned What LlamaCon announced, done today. A then-and-now lesson: a playground that sends one question to two Muse models side by side, a Llama 3.2 3B tune with a blind check in place of the Llama API's 8B tuning, and a link to the protections lesson.

• Muse Wonder
• Spark 1.3 Contributor

Tune: Llama 3.2 3B LoRA on a Lambda A10, about $2 to $3.

Estimated cost to redo: About $4 to $12 in all. Nothing is spent until the owner says so.

Not checked yet: What the keynote itself covers; Build cost of the playground; A tune launched through the site.

Muse Muse today.

Nothing built for this lesson yet.

Planned Fine-tune faster with Unsloth, measured. Run the Unsloth QLoRA Muse Muse already uses on Llama 3.2 3B, then Glimmer 30B, and show peak memory, time and cost next to Unsloth's claims, which stay labeled as theirs.

• Spark 1.3 Contributor
• Muse Wonder

Tune: Llama 3.2 3B on a Lambda A10, then Muse Glimmer 30B on an H100.

Estimated cost to redo: About $2.15 for the 3B; about $2.40 more for Glimmer. Nothing is spent until the owner says so.

Not checked yet: Any speedup measured by us (needs a plain baseline run); Llama 4 Scout as a base (never run here).

Muse Muse today.

Nothing built for this lesson yet.

Planned Run your own Llama on Linux, and keep answering from it. Tune Llama 3.2 3B, take its 2 GB Q4_K_M file, serve it with llama-server on Linux, and build a receipt reader that falls back to it when the cloud is cut, shown in an outage drill.

• Muse Wonder

Tune: Llama 3.2 3B LoRA on a Lambda A10, about $4.50 with 100 examples.

Estimated cost to redo: About $4.50 for the tune; $2 to $5 for the product. Nothing is spent until the owner says so.

Not checked yet: The Linux export script on a real Linux PC; Edge builds in a real sandbox; Whether Meta's old download flow still works.

Muse Muse today.

Nothing built for this lesson yet.

Planned Run your own tuned model on your Mac. The video's four steps on a model the learner tuned: a Llama 3.2 3B LoRA, its 2 GB Q4_K_M file in Ollama, the same curl and Python calls, and the stock model side by side.

• Muse Wonder

Tune: Llama 3.2 3B LoRA on a Lambda A10.

Estimated cost to redo: About $2.20 to $4 for the tune; $1.50 to $5 for the optional app. Nothing is spent until the owner says so.

Not checked yet: No tuned file has run on a Mac yet; The small-base local files path is off (QUANT_TARGETS); A page calling localhost:11434.

Muse Muse today.

Nothing built for this lesson yet.

Planned Own your model: open weights against a cloud API, tried for real. Do the talk's one hands-on idea: a Muse model writes the data, Llama 3.2 3B is tuned on an A10, a blind check scores it against its base, and you keep the file.

• Muse Wonder
• Spark 1.3 Contributor

Tune: Llama 3.2 3B LoRA on a Lambda A10; Muse Glimmer 30B on an H100 as an option.

Estimated cost to redo: About $2.15 per practice run; $17 to $36 with a 1,000-example Muse Wonder dataset. Nothing is spent until the owner says so.

Not checked yet: The talk's case-study numbers (no transcript); A 1,000-example Muse Wonder dataset; Training on Muse output (DATA_RIGHTS).

Muse Muse today.

Nothing built for this lesson yet.

Planned Tokenizers, Llama 3 to Muse. A concept lesson with the tokenizer facts and bugs Muse Muse logged in its own lab, plus a token counter that counts the same text on two Muse models through Meta's count endpoint.

• Muse Wonder
• Spark 1.3 Contributor

No fine-tune: the lesson has none.

Estimated cost to redo: About $2 to $5 to build; no GPU. Nothing is spent until the owner says so.

Not checked yet: Whether a built product may call the token counting endpoint; Whether that endpoint is free.

Muse Muse today.

Nothing built for this lesson yet.

Planned Where your model can run. A tour of where a Muse Muse model runs, each stop labeled by how far it has really run: Muse Wonder through the API, your own tuned 3B in llama.cpp or Ollama, a Jetson board fallback, and hosted GPUs marked not yet tried.

• Muse Wonder

Tune: Optional: Llama 3.2 3B LoRA on a Lambda A10, about $1.50.

Estimated cost to redo: About $0.05 for the receipt reader build; about $1.50 for the optional tune. Nothing is spent until the owner says so.

Not checked yet: Hosted GPU routes against a real provider; A tuned Llama file on a Jetson; The export scripts with a tuned file.

Muse Muse today.

Nothing built for this lesson yet.

Planned Distill a big model into a small one. Muse Wonder writes the data, a Llama 3.2 3B student trains on one A10 with only a one-line prompt, a blind judge compares it with the untrained student, and a second round shows where more synthetic data stops helping.

• Muse Wonder

Tune: Llama 3.2 3B LoRA on a Lambda A10; Muse Glimmer 30B on an H100 as the larger student.

Estimated cost to redo: About $4 to $9 for a 200-example run. Nothing is spent until the owner says so.

Not checked yet: Training on Muse Wonder output (DATA_RIGHTS Q2); A Llama student through the platform's own launch.

Muse Muse today.

Nothing built for this lesson yet.

Planned Watch guide: AI and the next platforms. A watch guide that links the talk, then three try-its that point at runs Muse Muse already logged: a Llama 3.2 3B tune you keep, a chat helper for a made-up business labeled as AI, and the edge track.

• Muse Wonder
• SAM 3.1

Tune: Llama 3.2 3B LoRA on a Lambda A10, about $2.15.

Estimated cost to redo: $0 for the page; about $2.15 for a tune; $3 to $8 for the optional build. Nothing is spent until the owner says so.

Not checked yet: Chapter times and quotes (no transcript); Build cost for this muse.

Muse Muse today.

Nothing built for this lesson yet.

Planned Run your own tuned Llama on a Windows PC. Tune Llama 3.2 3B, make its 2 GB Q4_K_M file, run it on Windows with the generated PowerShell script, and ask the video's prompt to base and tuned side by side. The script has never run on Windows.

• Muse Wonder

Tune: Llama 3.2 3B LoRA on a Lambda A10.

Estimated cost to redo: About $4 to $10 in all. Nothing is spent until the owner says so.

Not checked yet: The Windows script on a Windows PC; The Ollama path; Speed on Windows hardware.

Muse Muse today.

Nothing built for this lesson yet.

Planned A docs agent that looks things up before it answers. Split docs into chunks with page numbers, give Muse Wonder a keyword search tool it calls before answering, quote the passage and page for every answer, and check it against no docs and all docs.

• Muse Wonder

Tune: Optional: a Llama 3.2 3B LoRA that answers only from given passages, on a Lambda A10, about $2 to $4.

Estimated cost to redo: About $2 to $5 to build; $2 to $4 more for the optional tune. Nothing is spent until the owner says so.

Not checked yet: The video's model, provider and code; Whether the tune beats its base.

Muse Muse today.

Nothing built for this lesson yet.

Planned Safety as a system: checks around the model. A lesson with one diagram, and a Muse Wonder lab that checks prompts, replies and code against a written hazard list, flags instructions hidden in pasted pages, and counts catches, misses and false flags on 40 made-up prompts.

• Muse Wonder

No fine-tune: the lesson has none.

Estimated cost to redo: About $2 to $5 to build; $0.50 to $1.50 per 40-prompt bench run. Nothing is spent until the owner says so.

Not checked yet: Build and bench cost; Any accuracy claim.