The Open Bench

Kansas farmers are selling beef straight to families instead of through the packers. Congregations are losing the priest who stood between a person and the divine. A writer reaches readers without a publishing house deciding first. The middle is what goes. The two ends, the person who makes and the person who uses, are what remain.

That is the tradesman's world coming back. The family business. The one whose customers know eir face. If you have a trade, you are better placed than the office worker who used to look down on you.

Here is the rule I run my own life on, and it is the first thing I would hand anybody starting out. If a machine can do it, it gets cheap, so give it away or stop competing there. If it takes your hands, your presence, or your name on the line, it keeps a price.

I sell tools I made at a bench from wood I cut. I give the writing away. Not as a gesture. As arithmetic.

My position, stated as mine: the inherited side of culture is about to boom, and it will boom because of the machines rather than in spite of them. Once the work that was only ever about staying afloat gets taken, people can turn to creative pursuits and to the inherited trades their circumstances used to price them out of. Artisan work, blacksmithing, IPA and mead, knitting, crochet, more and more of it. What it expands is the sheer number of experiences one person can have in a life.

Some of that is already measured.

The National Endowment for the Arts participation survey, which the Census Bureau fields and which reached nearly 41,000 adults, found the share of Americans doing leather, wood or metal work rose from 6.6 percent to 8.9 percent between 2017 and 2022. About a third more people in five years.

Enrolment at vocationally focused public two-year colleges reached 871,000 in spring 2025, up 19.4 percent across five years, and rose again this spring by 2.8 percent, another twenty-four thousand students.

Registered apprenticeships stood at roughly 680,000 in the 2024 fiscal year, up 114 percent on the decade before. And there were about sixty commercial meaderies in this country in 2003. There were more than five hundred in 2025.

I hold as a position that creation and exploration are what remain for us. They are the only two activities with no prior answer to correlate against. Everything that has a prior answer, the machine will find faster than I can. What is left is exactly the part I wanted: the object that did not exist, the idea nobody had, the place nobody has stood. I am not diminished by handing over the part with an answer key. I am freed for the part without one.

The purpose of life is to tell all the stories, and yours is one of them. There is no bottom to that. The stories that can be lived, made and explored are endless, and a machine that handles the paperwork does not use them up. It hands them back. Nobody who has ever been handed a free week said there was nothing left to do with it.

What decides whether a country full of automated work falls apart is not whether the machines arrive. It is who owns them and where the money goes.

Chesterton and Belloc argued a century ago that the problem was never the machines but ownership gathered into too few hands, and that the cure was property spread wide: a shop, a farm, a trade, a share, for as many households as possible. They called it distributism. A machine in a workshop you own is a tool. The same machine owned by a company you rent from is a boss.

That is why the free and locally run options come first in everything below. Refusing the tools does not get you out of the corporation's reach. Learning them well enough to run your own does. The door out already exists: weights you can hold, on hardware you own, in a room you are standing in.

People picture catching up as a fixed distance. Wait a year, spend a month learning, arrive where everyone else is. It does not work that way.

The skill is not the button. Anyone can type a question into a box. That was never the skill. The skill is decomposition, breaking a thing you want into pieces a machine can do well; direction, knowing how to specify so the shape stays yours; checking, knowing exactly where these systems fail and building the verification into the process; and workflow, the prompts, the memory files, the attribution habit, the corrections log, the pipeline from raw material to finished thing. None of it comes in the box. All of it is built by use, and all of it compounds.

The material accumulates. After a year you have a corpus: your own work, structured in forms the machines can use; your prompts, refined; your rules written down; a memory of your projects that survives between sessions. Whoever starts today starts with none of that, while everyone who did not wait has years of it. The tools reward whoever has been feeding them, and the reward grows with the feeding.

The floor rises around you. When one maker in a field can check forty sources and publish a corrections log, that becomes what a careful piece of work looks like. When one shop can list, photograph, tag and track without hiring, that becomes the expected pace. Your standard has not fallen. The floor has come up past it, and it keeps coming.

One thing does run the other way and I will not leave it out. The tools get easier. But what gets cheaper is the first hour. What compounds is the thousandth.

Refusal is usually argued as if it were free, as if the person who declines the tools keeps everything they had and merely forgoes a convenience. That is not what happens.

You do not get the same work, slower. You get less work. The honest counterfactual to a research pass across forty sources is not the same pass by hand. It is no pass. The counterfactual to the corrections log is no log. The counterfactual to the piece that shipped is the piece in the drawer.

You lose the check, and you do not notice losing it, because you do not experience your own error rate. It is invisible to you by construction.

You lose the room. The terms these systems get used on are being set now, by the people present. Anyone who refuses to touch this on principle will not slow it by a week. They will only be absent when the terms are written.

And you lose the collaborators. The scarcest thing in my creative life is other makers who are not paralysed by this. The hardest part of building a creative community right now is finding artists who are not frightened of this and fighting it. The fear is understandable and some of it is correct. But it leaves a lot of good people with no plan except refusal, and refusal is not a plan.

The trades already ran this experiment. The makers who refused the machine tools of the last century were not remembered as the purists. They were not remembered. The ones we remember took the new tools into the shop and put their hands back where hands are irreplaceable. That is what I am doing. The machine gets the answer key. I keep the wood, the fire, the stone and the sentence.

Everything I write is free, on every platform I can reach, in every format I can produce. What I sell comes out of my hands or my voice: the tools I make at the bench from wood I cut and seasoned and stone I set by hand, with the audiobooks I read aloud, the handmade editions and the hardbacks I sign all on their way. If the machine can do it, I give it away. If it takes my hands, it has a price.

Draw that line first and most of the rest of the business follows from it.

Every finished piece carries a note in three parts. Direct input is my own writing and thinking, carried as written. Directed or requested is work I specified closely enough that the shape is mine. AI researched or generated is work produced without line-level direction from me, which I then checked. The three add to a hundred, and the third number is never rounded down to look better.

Every finished piece. Not the ones where the split happens to flatter me.

Where a claim was too strong, it goes in that piece's corrections log rather than being quietly fixed. Where sources conflict, the conflict is stated rather than resolved.

Sixteen entries on one essay, every one removing a claim that was too strong. That is not the machine making me sloppier. That is the machine making me accountable in public, and it has improved the writing every single time.

A corrections log is not a nice habit. It is the minimum condition under which anyone should be allowed to use these systems on anybody else.

Three options for any claim: print it if it is supported, leave it out silently if it is not, or state it plainly as a position in your own voice. Never print a thing and then knock it down in the same breath.

A folk figure you cannot source does not get a hedge. It gets killed before it reaches the reader.

I run my work past more than one model and ask each to find what is wrong with it, and I tell it to be rude. The fact checks have caught fabricated citations and scholars quoted backwards. They have also caught me.

A tool that only agreed would be worthless to me, and the people who say these systems only flatter have not asked them to do anything else. I would rather be corrected than agreed with.

Both things are true, and which one happens to you is decided by what you hand over.

A person who hands over their thinking will get lazier, and will not notice, because the output stays good while the ability underneath it goes. A person who hands over the parts that were never the work, the formatting, the sorting, the second draft of the same paragraph, the thing that took two hours and taught nothing, gets those hours back and spends them on the part that is hard. That person gets sharper, because the hard part is now the only part left, and they are doing more of it.

The tool does not decide which of those you are. You do, every time you type into it.

I have AuDHD. The gap between having something to say and having it in shippable shape has swallowed years of my output, not because the thinking was missing but because the setup work, the formatting, the fortieth small decision before publishing, is where my kind of mind stalls.

My standing instruction to myself is to ship unfinished rather than wait for ready, and these tools are most of what makes that possible. They hold the thread when I drop it. They remember the state of a project across the weeks I am away from it. There is writing of mine in the world today that would otherwise be in a drawer, and I do not mean a little of it.

The version of trust I prefer needs no biometrics: verification guilds. Makers staking their own reputation to vouch for each other, in public, and able to withdraw it.

It scales badly and it can be captured. Both are real objections. It is also how every craft guild in history solved this, and it puts the judgment with people who have something to lose.

The objection about depending on a company is fair today, and the door out already exists. Anyone who wants these capabilities without a landlord can have them now, at a lag from the frontier.

Where you cannot own it, know what you are renting and what it would cost you to leave.

Rule six is the best supported claim on this page. Five studies across five different methods converge on it: hand over the reasoning and measured ability drops, hand over the routine and output rises with no measured skill cost.

The sharpest single finding is that both modes deliver comparable immediate benefits, so a person may not be able to tell from experience which one they are in. That part is a self-reported association rather than a measured one. The measured half comes from the Turkish school experiment, where the students given an unrestricted chatbot scored below the students who never had one at all once the machines were taken out of the room. Between them, that is why this has to be a rule rather than an instinct.

Sources: Zhu and colleagues, Frontiers in Psychology, 16 July 2026, n=589. Bastani and colleagues, PNAS, 2025. Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), April 2025, n=5,172. Dell'Acqua and colleagues, Organization Science, 11 March 2026, n=758. Lee and colleagues, CHI 2025, n=319.

The exit from the subscription. Free, offline, and nothing leaves the room.

Hardware, honestly: 8 GB of RAM runs 3 to 4 billion parameter models, which is fine for tagging and reformatting and not fine for drafting. 16 GB of unified memory, or 8 to 12 GB of VRAM, is the realistic entry point. 32 GB unified or 24 GB VRAM gets you the 26 to 31 billion class, which is where a local model starts feeling like a real assistant.

• Ollama. One command to install, one to run a model. The easiest on-ramp. Free for local use.
• LM Studio. A desktop app if you would rather not touch a terminal. Free tier.
• Jan. Open source, chat-shaped, local or bring your own key. Free.
• llama.cpp. The engine most of the above are built on. Runs on almost anything. MIT licence.
• Open WebUI. Self-hosted web front end that makes a local install feel like a hosted product. Free.
• Qwen3.8-27B and Gemma-4-26B-A4B. The two best licensed models for commercial work right now, both Apache 2.0. Gemma's mixture-of-experts design activates only about 3.8 billion parameters per token.
• Hugging Face. Where the open models actually live. Free public hosting.

Where the drafting, the structural editing and the verification passes happen. All four assistants below have a real free tier.

• Claude. Longform composition, structural editing and verification passes. Free tier covers web, mobile and desktop with search and file creation.
• ChatGPT. Fast ideation and brainstorming. Free tier gives unlimited text chats on the current default model.
• Gemini. The most generous free tier of the three, and it reaches into Gmail and Drive.
• Mistral. European, and the free tier includes ten dollars a month of API credit if you want to build something small of your own.
• Zotero. Free, nonprofit reference manager, and the direct antidote to a fabricated citation. If it is not in Zotero with a real link, it does not go in the piece.

Listing photos, social images, logos, cut files. Licence detail matters more here than anywhere else on this page.

• GIMP, Krita and Inkscape. Free and open photo editing, painting and vector work. The GPL covers the software, not your artwork. You own and may sell what you make.
• Photopea. Browser-based Photoshop clone handling 40-plus formats including PSD, AI, PDF and RAW. Runs client-side, so nothing is uploaded. Free.
• ComfyUI and Fooocus. Local image generation, from most control to least setup. Free and open source.
• FLUX.1-schnell. Apache 2.0. Model and outputs both free for commercial use, which makes it the safest licence in local image generation.
• Upscayl. Offline upscaling, AGPL-3.0. Needs a Vulkan-capable GPU, and most integrated graphics will not run it.
• Canva. Fastest route to listing images and social templates. Free tier includes 1.6 million templates and 5 GB of storage, capped at 20 AI uses.

A complete zero-cost pipeline exists here. This is the section where free genuinely matches paid.

The standout is Kdenlive's built-in speech to text, which ships with two local engines, VOSK and Whisper. Mark a zone, pick an engine, click process. That is a full captioning workflow inside one free application, running entirely on your own machine.

• DaVinci Resolve. The free version is a real professional editor, not a trial: full edit, colour and audio up to UHD 4K at 60fps. Studio is a one-off 295 dollars.
• Kdenlive. GPL editor with local speech to text built in. Free.
• Shotcut. Lighter still, no import step. Free and open source.
• Audacity. Recording and cleanup for voiceover. Free and open source.
• OBS Studio. Screen recording and live streaming. Free and open source.
• whisper.cpp. Local transcription that runs fast on a plain laptop with no Python. MIT licence.

A complete, self-owned, zero-subscription back office. Every component below is free, and together they are the most concrete version of rule nine on this page.

• GnuCash. Proper double-entry bookkeeping on your own machine, with explicit small-business support. GPL.
• Wave. Free cloud invoicing and bookkeeping: unlimited invoices, estimates, bills and a mobile app. You pay only card processing at 2.9 percent plus 60 cents.
• Invoice Ninja. Invoicing, quotes, time tracking, expenses and a client portal. Self-host free, or five clients free in the cloud.
• Nextcloud. Your own files, calendar and office suite, self-hosted, running on hardware as small as a Raspberry Pi. Free.
• LibreOffice. Free office suite for personal and business use, no subscription. MPL v2.0.
• WooCommerce. A self-hosted store you actually own. Core plugin free and open source; the real costs are hosting and payment fees.
• Payhip. Sell digital goods with no monthly fee at all: zero dollars a month plus 5 percent per transaction, unlimited products.

A record of origin, in the order you can actually adopt it. Rule eight is the human half of this. The tools below are the technical half.

Worth knowing before you start: C2PA signs with ordinary certificates rather than a blockchain, and it says so explicitly. Anchoring those signatures to a chain is a separate layer, and it is what makes a record outlive the certificate authority. Also worth knowing: most platforms strip metadata on upload, which is why the durable-watermark and on-chain approaches exist at all.

• Adobe Content Authenticity. The realistic pick if you are not technical. Batch-sign up to 50 JPG or PNG files at once, attach a verified name and social accounts, and set a generative-AI training preference. Free, no paid subscription needed, still in beta.
• Verify. Paste a file or a URL and read its credentials, signer, timestamps and AI assertions. Free.
• C2PA. The open standard itself, steered by Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI, Sony and TikTok among others.
• IPTC Digital Source Type. The free controlled vocabulary for saying honestly how a file was made, including humanEdited and compositeWithTrainedAlgorithmicMedia. CC BY 4.0, usable today, costs nothing.
• CAWG. Extends C2PA with the assertions a solo maker actually wants, including identity and a do-not-train-on-this assertion.
• c2pa-rs and c2pa-python. The command-line and Python routes to signing your own files. Note the old c2patool repository was archived in December 2024, and many guides still link the dead one.
• ERC-7053, the Ethereum standard for recording media history on chain, status Final. Used by Capture, which registers provenance at the moment of capture so the record survives metadata stripping. Free to start.

FLUX.1-dev is not the same licence as FLUX.1-schnell. Under section 2(b) you may only use the dev model itself for non-commercial purposes. Under section 2(d) you may use the output for any purpose, commercial included. Plain version: you may sell the pictures, you may not build a commercial service around the model. If that makes you nervous, use schnell, which is Apache 2.0 end to end.

The default background-removal model in rembg requires a paid agreement for commercial use. The rembg tool itself is MIT, but the model weights carry their own licences, and the one it reaches for by default is not free for business. Switch it to u2net or another commercially cleared model before you put a single listing photo through it.

Craft fair juries commonly ban AI-generated work outright. Crafty Wonderland in Portland, to take one that publishes its rules, excludes mass-produced, imported, MLM and AI-generated art, while allowing custom-printed items using your own designs. Using these tools for admin, pricing, research, copy and bookkeeping is uncontroversial. Using them in the object you are selling is where applications get rejected. Read each jury's rules before you apply, not after.

The FTC's Consumer Reviews and Testimonials Rule, 16 CFR Part 465, has been in force since 21 October 2024, with civil penalties available for knowing violations. It bans fake reviews, incentives conditioned on positive sentiment, undisclosed insider reviews and bought followers. Asking real customers for honest reviews is fine. Generating them is not.

Fabricated citations are measured, not folklore, and the rate is rising. One study of ChatGPT bibliographies found 55 percent of GPT-3.5 citations and 18 percent of GPT-4 citations were fabricated. A 2026 analysis of 111 million references across 2.5 million papers found at least 146,932 citations in 2025 papers pointing at sources that do not exist, and that 85.3 percent of the hallucinations present in preprints persisted into the published record. Check every reference against the source. That is rule three in practice.

Do not cite the g equals 0.867 meta-analysis on ChatGPT and learning. It was published in May 2025 and retracted on 22 April 2026 over discrepancies in the analysis, with the authors not responding to correspondence. It is still the single most-quoted figure in optimistic writing about AI and learning, and it is dead.

Four of these are paid and two have a real free route. The licence line matters more than the price line.

• Midjourney. The aesthetic benchmark. Current default is V8.2, from 24 July 2026. Basic is 10 dollars a month, there is no free tier and no trial, and below the Pro plan everything you generate is public by default, which surprises people.
• FLUX. Black Forest Labs, now on FLUX 3 with the FLUX.2 family beside it. Read the licence per model: FLUX.2-klein-4B is Apache 2.0 and free for commercial use, while klein-9B and FLUX.2-dev are non-commercial despite being freely downloadable.
• Stable Diffusion. Still on the 3.5 family; Stability shipped no new image foundation model in 2026 and spent the year on audio instead. Its Community Licence is free including commercial use until you pass a million dollars of revenue.
• Nano Banana 2 Lite. Google's fast image model, four seconds to an image, with strong character consistency for storyboarding. About 3.4 cents an image through the API, which excludes image models from its free tier. Free in the Gemini app, but app images carry a visible AI marker as well as an invisible watermark.
• Adobe Firefly and Photoshop Generative Fill. Adobe states its outputs are safe for commercial use, which is the reason to pay for it. A genuine free tier gives limited generations that refresh daily; Standard is 9.99 a month.
• Luminar, which has dropped the Neo from its name. Sky replacement, portrait and skin tools, RAW for a thousand-plus cameras. 129 dollars once, seven-day trial, no card. Read this part before buying: the one-time price covers the base app, and the AI tools run for a year before they need a subscription to renew.

The free item in this list is the surprise, and it is the one nobody mentions.

• Roland Melody Flip. Free with any Roland Cloud membership including the free tier, which makes it the best no-money tool in this section. Feed it your own track, loop or beat; it reads structure, tempo, key, chords, genre and mood and hands back melody, chord, bass and drum parts. VST3 and AU, so it runs inside your own DAW, and it exports MIDI as well as audio. Built with Sony Computer Science Laboratories.
• Suno Studio 2.0, from 13 August 2026. A browser digital audio workstation with AI in it: MIDI import, recording and editing on a timeline, controller support, stem separation into vocals, drums and bass, and a chat bar you talk to in order to build instruments and plugins from scratch. It needs the Premier plan at 24 dollars a month. The free Suno plan carries no commercial rights at all, which is the single most common trap in AI music.
• Google Flow Music, running the Lyria 3.5 model. Generate, edit and extend songs, with a real free tier: free signup, daily top-up credits, all core features, two generations at a time. Requires a Google account, 18 or over, in a supported region. Google does not publish commercial terms for the output, so do not assume them.

Both of these are paid, and both landed within the last few months.

• Premiere Pro and After Effects Generative Media, announced 8 September 2026. Generates context-aware, fully editable video and audio clips directly in your timeline, using reference frames from your own project so the result matches the surrounding edit. Four tools: Generate Video, Generate Sound Effects (you can conduct the rhythm with your own voice), Generate Soundscape and Generate Music, the last of which Adobe describes as commercially safe with synchronisation rights. It is model-agnostic, so Firefly, Veo, Kling, Runway and Luma are all selectable. Premiere is 22.99 a month on its own.
• Google Vids with Gemini Omni. Make and edit video by describing what you want, and build a personal avatar from a selfie and a short voice recording that then delivers what you type. Paid only, through Google AI Pro or Ultra or a Workspace business plan. Avatars are 18 and over and region-limited, and everything carries a SynthID watermark.

Two of these have real free tiers and the two fiction tools do not, though one of them lets you supply your own free model.

• Sudowrite. A fiction specialist. Its Story Bible carries characters, world and outline forward so a long manuscript stays consistent, which is the thing general assistants are worst at. From 10 dollars a month; free trial, no permanent free tier.
• novelcrafter. A novel workspace where you bring your own API key, and the Codex, its worldbuilding store, can be shared across a whole series. Worth knowing for this page: it connects to LM Studio and Ollama, so the model can be one running on your own machine for nothing, and the four to twenty dollars a month buys the workspace rather than the words.
• ProWritingAid. Free tier is real but capped at 500 words a document, which is a scene rather than a manuscript. Premium is 30 a month or 90 a year, with a lifetime option.
• Grammarly. The most generous free tier in this category: grammar, spelling, tone and 100 AI prompts a month at no cost.

Two of the four best-known names here are research papers rather than things you can download, and the list says so rather than letting you find out after an hour of searching.

• Summer Engine. A complete game editor that opens existing Godot 4 projects unchanged and hands them back to stock Godot with no lock-in, plus an asset store. The editor, manual editing, native export and the MCP server are free forever; sustained AI generation needs credits. The free part is a real engine, not a crippled demo, which makes this the strongest entry here for somebody with no money.
• Unity AI. Generators inside the Unity Editor for sprites, textures, sounds, animations and materials from text prompts or reference inputs. Note that 3D model generation is not among them, whatever you read elsewhere. Fourteen-day trial with 1,000 credits, then 10 dollars a month. Generated assets are stamped with metadata marking them AI-generated, and Unity puts the app-store declaration duty on you.
• WorldGen, Meta Reality Labs. Generates traversable, fully textured 3D environments from text that can be explored or edited in a standard game engine. It is a research project with a paper, not a product: no download, no signup, no price. Worth watching, not worth planning around. Note also an unrelated open-source project of the same name, which is a different thing entirely.
• Hunyuan3D-WorldClaw, Tencent Hunyuan, August 2026. An agentic framework that turns a single open-ended prompt into an explorable, editable open-world 3D scene. Also research. A repository exists but the licence could not be confirmed, and Tencent's earlier 3D models shipped under terms with regional and commercial restrictions, so read the licence file before building anything on it.

The best item here is free, MIT licensed, and came out of a peer-reviewed conference rather than a launch.

• Open-Theatre. An open-source toolkit for LLM-based interactive drama: character agents, memory, and four multi-agent coordination strategies including director-actor arrangements. Published as an EMNLP 2025 system demonstration. MIT licence, runs locally with your own keys, so with a local model it costs nothing to run and carries no commercial restriction.
• Figma. Config 2026 added Figma Motion, a real timeline with keyframes that exports to CSS, JSON, React, MP4, WebM, SVG and GIF; AI-generated shaders you describe or feed a reference image; and Code Layers, which turn any layer into interactive code and back. The free Starter plan is three projects and two editors. Figma does not publish which tiers get the new AI features, so check before you plan around them.
• Cursor. The AI code editor, useful pointed at a Godot or Unity project while the engine handles running it. The Hobby tier is free with limited agent requests; Pro is 20 dollars a month.

One genuine bargain here, and one of the worst free-tier traps on the whole page.

• Marvelous Designer. Garment simulation and pattern-based clothing, from CLO Virtual Fashion. 39 dollars a month, or 280 prepaid for a year, and 8.25 a month on a student licence with a school email, which is the cheapest serious garment simulator available. Fourteen-day trial. Standalone perpetual licences ended on 2 December 2025, so it is subscription only now.
• CLO-SET, the collaboration platform from the same company, has a genuinely free tier: three users, ten file uploads per group, 50 MB of rendering a month. It is a different product from CLO 3D itself, which does not publish prices where they can be read.
• Tripo AI. 3D models from text or images, exporting GLB, FBX, OBJ and USD into Blender, Unity and Unreal. The free tier gives 200 credits, about thirteen models, and it is labelled public models and non-commercial use. Anything you make on the free plan is visible to everyone and cannot be sold. Commercial rights start at 20 dollars a month.
• Canva, now organised as Standard, Premium and Ultra AI tiers rather than the older Magic Suite branding. The free plan includes up to 20 AI uses, and Magic Resize, Magic Translate and the background remover are all behind Pro.

The free and local option is the one worth starting with.

• Podcastfy. Apache 2.0, free and open source, and it runs locally against your own models, including local ones, so nothing leaves your machine. Turns documents, links and notes into a scripted, multi-voice episode.
• PODLM. Turns URLs, text and documents into podcasts with customisable multi-speaker voices. Basic is 6.99 a month billed yearly, Pro 11.99. There is a trial rather than a free tier, and commercial terms are not addressed.
• Rebel Audio. An all-in-one platform aimed at first-time creators: record, and it edits, cuts social clips, writes titles and descriptions, makes cover art and publishes to Spotify, Apple Podcasts and YouTube, with voice translation into more than 47 languages. Public since 30 May 2026. No free tier, and pricing is behind a signup.

One of these is free and better than the paid tools it competes with.

• NotebookLM, which Google is in the middle of rebranding as Gemini Notebook. Upload your own sources and it answers from them with citations, plus audio and video overviews, reports, flashcards, quizzes and mind maps. Free gets 100 notebooks, 50 sources each, 50 chats a day, three audio and three video overviews a day, and ten deep-research runs a month. If you only take one thing from this section, take this one. Note that its address is moving from notebooklm.google to notebook.google, so old links will rot.
• Gamma. Generates presentations, sites and documents from a prompt. Be clear about the free tier: it is 400 credits that never refresh rather than a monthly allowance, output carries Gamma branding, and prompts are capped at ten slides. It is an extended trial wearing a free tier's clothes.
• Beautiful.ai. Smart templates that re-lay themselves out as content changes. Pro is 14.50 a month billed annually. There is no free tier, and the fourteen-day trial requires a card that charges automatically if you forget to cancel.

Start with the free ones. There are more of them than the paid lists admit.

One safety note before the list. The AI Hub community's vanity address, discord.gg/aihub, was hijacked on 30 April 2026 and now points at a malicious clone. Its own documentation says so. Any list still printing that address is sending people somewhere dangerous. The official invite is the one below.

• AI Hub Discord, 517,000 members, free. Centre of gravity is voice models and AI vocal covers, alongside art and tooling. Use this invite and not the old vanity link.
• Civitai. The big model and LoRA community. Two things people get wrong: content reproducing the likeness of real people is prohibited outright, and licences are set per model rather than site-wide, so a maker selling work has to read each model's own terms.
• SeaArt. Image, video and audio community with a free tier. Its own February 2026 release reports over 50 million registered users and more than 30 million monthly active, which is a company figure rather than an audited one.
• Suno Discord, 395,000 members, free, and Suno's own free tier gives ten songs a day without a card, though commercial rights attach only to the paid plans.
• Hugging Face Discord, 237,000 members, plus the forums and Spaces, where the community's own image, video, voice, music and 3D apps run free on shared hardware.
• Ollama's Discord (197,000 members) and ComfyUI's (61,000 members). Free and open source, and between them the practical route to generating text and images on your own hardware at no cost per generation.
• Adafruit's Discord, 39,000 members, free. A long-running maker and electronics community where AI tooling is discussed openly rather than banned, which is rarer than it should be. Indie Hackers is the equivalent for people building a small business.
• Free courses from real institutions: the Hugging Face diffusion models course and its audio course; fast.ai Practical Deep Learning for Coders, nine lessons recorded at the University of Queensland, which assumes about a year of coding and no deep learning; Elements of AI from the University of Helsinki, which assumes neither; Microsoft's Generative AI for Beginners, 21 lessons, MIT licensed, though its exercises assume an API key unless you run models locally; and Claude Academy, which is framework-led and aimed at people who do not code. Disclosure, since this section is written by Claude: that last one is Anthropic's.
• Goldsmiths MSc Creative AI. Fully online, two years part-time, 180 credits, 12,500 pounds total or paid per module. Entry is a 2:1 in any subject and non-technical backgrounds are explicitly welcomed. Four intakes a year; the next starts 26 October 2026 with applications closing 2 October. Goldsmiths also runs a campus-based MA Computational Arts at 14,300 pounds full-time for UK students, with no prior coding required.
• The UAL Level 3 Extended Diploma in Creative Media Production and Technology is worth understanding properly, because it is widely misdescribed. UAL Awarding Body publishes the qualification and approves colleges to deliver it; it does not teach it. The AI and generative media route is a college's own curriculum choice, for example at Capital City College, where it covers RunwayML, Midjourney, DALL-E and Firefly, and is free for students who were 16 to 18 on 31 August 2025. One time-sensitive fact: the Extended Diploma is defunded in England from autumn 2027.

Every one of these is a free tier or a headline price that costs somebody something they did not expect. They are gathered here so they are hard to miss.

• Suno's free plan carries no commercial rights. Music you make on it cannot be sold or used in anything you sell.
• Tripo AI's free tier is non-commercial and makes your models public. Assets generated on it cannot ship in a commercial game.
• FLUX.2-klein-9B and FLUX.2-dev are non-commercial licences on freely downloadable models. klein-4B is the Apache 2.0 one.
• Civitai licences are per model, not site-wide. Read each model's own terms before selling work made with it.
• Unity AI stamps generated assets with AI metadata and makes the app-store declaration your responsibility.
• Beautiful.ai's trial requires a card and charges automatically. Gamma's free credits never refresh and its output carries Gamma branding.
• Luminar's one-time price covers the app; the AI tools stop after a year without a subscription.
• Midjourney generations are public below the Pro plan.
• Gemini app images carry a visible AI marker as well as an invisible one.
• The discord.gg/aihub address was hijacked in April 2026. Use the invite listed above.

This list arrived with four entries that could not be verified, and naming them is more useful than deleting them, because the same four are circulating in every similar list.

Podify, described as a local AI podcast studio that researches, scripts and generates a finished episode. No such software was found. The domain belongs to a human video-podcast production agency in Los Angeles whose flagship package is 5,500 dollars. The tool the description is reaching for is almost certainly Podcastfy, which is open source, free, runs locally, and is listed above.

SceneStudio, described as a tool for cinematic visual novels with AI scripts, character art and video scenes. The nearest real product with that name is a static image generator that does not write scripts and does not generate video.

Nessie, described as a research assistant answering from your own notes and PDFs with citations. Nessie is real but does something else entirely: it indexes your AI chat history across tools so you can search and resume sessions. It does not read your document library.

Chunk, described the same way. It is a real iOS app, but it sells weekly subscriptions and its own user reviews report the document feature failing to see uploaded documents. NotebookLM does what both of these were claimed to do, better, for nothing.

Two feature claims also went unsupported and were left out rather than repeated: that the current Midjourney renders five times faster with native 2K output, and that Google Flow Music splits stems. Neither appears in the vendor's own material.

None of this means the model that produced the list was careless. Eleven of fourteen items in one category checked out exactly, including two names that looked invented and were not. It means a list is a claim until somebody opens the links.

You do not need to own a shop. You need access to one, and in most cities that costs less per month than a phone bill, or nothing at all.

• Tool libraries are free. Portland runs six of them, and the city's own page says plainly that membership is free and everyone is welcome. Bring photo ID and proof of address and borrow up to seven tools a week.
• Guild makerspaces are the next step up. Past Lives Makerspace in Portland is 26,000 square feet organised into thirteen or more guilds, wood, metal, textiles, ceramics, glass, jewellery and writers among them, with 24/7 access at 180 to 200 dollars a month, and joining a guild costs nothing on top of membership. ADX Portland starts at 75 dollars a month.
• Elsewhere: Artisan's Asylum in Somerville runs 60 to 150 dollars a month with subsidised rates published for financial need, and Pumping Station: One in Chicago is 65 dollars a month with a zero-dollar six-month scholarship tier. NYC Resistor in Brooklyn runs free Craft Nights on Mondays and Thursdays.
• Visit a guild before you pay anything. Gathering of the Guilds brings nine Portland and Oregon guilds under one roof, metal arts, woodworkers, potters, glass, weavers, sculptors, printmakers, faceters and the bead society, and all of them hold recurring meetings that welcome visitors.
• Find one near you through the free Nation of Makers finder or the Make: directory.
• If nothing is near you, start one. Shareable's Library of Things Toolkit 2.0 is a free and genuinely detailed how-to covering inventory, space, governance, money and law, written for people starting on nothing.

Two routes cost nothing: get paid while you learn, or get a scholarship. Both are real and both are underused.

• Registered apprenticeship pays you a wage from day one. The Apprenticeship Job Finder searches openings by trade and location and lets you apply to the employer directly. Both Oregon and Kansas run their own state apprenticeship agencies, so you deal with the state rather than the federal office.
• The craft schools give away more than people realise. Penland publishes full scholarships covering tuition, housing and meals, alongside work-based awards, studio assistantships and a standby programme at half tuition, and says about half its workshop students receive some financial assistance. Summer 2027 scholarships open on 15 October 2026 and close on 12 November. The John C. Campbell Folk School runs Windgate scholarships covering tuition, housing and meals. Arrowmont runs paid work-study. Haystack selects for a free Open Studio Residency and gives a 10,000 dollar Artist Grant with a six-month mentorship.
• Blacksmithing specifically: ABANA is 35 dollars a year for a digital youth membership and 75 standard, and gets you a three-level national curriculum plus more than eighty local affiliate chapters, which is where the actual hammer time is. Peters Valley also runs blacksmithing as a core discipline.
• Lapidary: local clubs under the American Federation of Mineralogical Societies are the standard cheap route to cutting equipment and instruction. Oregon falls under the Northwest federation, Kansas under the Midwest.
• Community college is the priced middle. Portland Community College publishes eighteen welding and metal fabrication career certificates each under a year, at 143 dollars per credit plus about 11.60 in fees.
• Free and open: MIT OpenCourseWare carries materials from more than 2,500 courses with no registration, and a public library card gives you databases, computers, printing, bookable rooms and explicit small-business help.

Grants for individual makers are thinner than the internet suggests. Zero-interest lending and local arts money are where the real doors are.

• Kiva US lends 1,000 to 15,000 dollars at zero percent interest, with no fees, no collateral and no minimum credit score, on character-based underwriting, repaid over 12 to 36 months. It excludes Nevada and North Dakota and requires a PayPal account. Budget roughly two months from application to funded. This is the best-fit capital source on the page.
• Local arts money beats national arts money. Portland's RACC Arts Project Grant is open to individual artists in Multnomah, Clackamas and Washington counties, with the current window running 1 September to 28 October 2026; its 2026 second round put 1,220,000 dollars into 274 recipients, 179 of them individuals.
• The Awesome Foundation gives 1,000 dollars with no strings and no equity taken. The Portland chapter funds every other month and explicitly seeks individuals and groups without nonprofit status.
• FCA Emergency Grants awards 500 to 3,000 dollars on a rolling basis, roughly fifteen to twenty a month, for a sudden opportunity or unexpected costs on a near-complete project with committed dates.
• The honest answer on the NEA: its main programmes fund organisations, not individuals. The one real individual route is the 25,000 dollar Creative Writing Fellowship. Go to your state and city arts councils instead.
• Watch this date. CERF+, the resource every craft list leads with, has suspended all grantmaking for the remainder of FY2026, with programmes expected to resume after 1 October 2026. Do not plan around it before then.
• Free advice really is free. SCORE mentoring is free for the life of the business, unlimited sessions, matched within 48 hours, open to anyone with a US business or even just an idea. SBDCs advise free. An EIN from the IRS is free and instant, so never pay a third-party site for one.
• Forming the thing is cheap. In Oregon an assumed business name (PDF) is 50 dollars and an LLC is 100 to form plus 100 a year, so a sole proprietor trading under a business name can be legal for fifty dollars. Kansas publishes no fee figures; call the Secretary of State.

Know the real cut before you pick. Every figure below is as published on 10 September 2026.

• Etsy takes 20 cents per listing and 6.5 percent of the displayed price plus shipping and gift wrap, plus payment processing, plus 15 percent Offsite Ads under 10,000 dollars a year and 12 percent above. The shipping inclusion penalises heavy work like canes and ironwork more than the headline rate suggests.
• Artisans Cooperative is the artist-owned alternative, currently in Phase 2 beta. Buy-in is 1,000 points at a dollar a point with no annual dues, and points can be earned through member activity or paid in instalments, so it is not a hard cash wall.
• Faire for wholesale takes 15 percent on a reorder, 15 percent plus a 10 dollar fee on an opening order from a new customer, and nothing at all on Faire Direct orders you bring yourself. Bringing your own stockists turns it into free order management.
• Square Free has no monthly fee and takes 2.6 percent plus 15 cents in person. That plus a reader is the standard craft fair setup.
• For digital goods: Payhip is zero a month plus 5 percent, Ko-fi is 5 percent or 12 dollars a month for zero, and Gumroad is 10 percent plus 50 cents with no monthly fee.
• Markets: Portland Saturday Market accepts new vendor applications year-round on separate craft and food tracks, the craft one now taking cottage-food-exempt products too, and publishes no fees at all, so you have to call. Crafty Wonderland charges 10 dollars to apply, credited against a booth at 425 dollars for five by ten or 575 for ten by ten. Applications for the December 2026 market closed on 22 August and it takes no late entries under any circumstance, so the date to diarise is the opening of the next cycle rather than its deadline.
• Before you sell anything physical, read the FTC on Made in USA, the FTC Endorsement Guides, CPSC small-batch registration for anything intended for children, and the SBA on product liability. Oregon has no sales tax, but selling online can still create obligations in other states.

If the argument above about who owns the machines is right, then the ownership structure is not an afterthought. There is a whole apparatus for this and almost nobody starting out knows it exists.

• Democracy at Work Institute is the free first stop on zero budget: a resource library aimed at startups, a webinar series on democratic management, and free referrals to actual co-op developers.
• CooperationWorks! finds the developer nearest you, free, in all fifty states. In the Pacific Northwest that is the Northwest Cooperative Development Center, which serves Oregon, Washington, Idaho, Alaska and Hawaii, and runs a ten-week Co-op Academy free for rural residents and 98 dollars for urban.
• US Federation of Worker Cooperatives publishes its full dues schedule: 300 dollars a year for a startup, or 100 for the first year if you are receiving developer assistance. It runs a technical-assistance clinic and a worker health benefits programme.
• Shared Capital Cooperative is a CDFI that is itself a co-op owned by its borrowers, lending 5,000 to 1,000,000 dollars over one to ten years, startup loans included. You have to join before you can borrow.
• Start.coop runs a free eight-week incubator with a 1,000 dollar stipend, though eligibility is restricted to rural US entrepreneurs and specified founder groups, so check the map before you spend time on it. Applications are closed as this is written; it runs in cohorts.
• The Platform Cooperativism Consortium helps you start, grow or convert to a platform co-op. Grassroots Economic Organizing is the living movement press, free and actively publishing. Cooperation Jackson in Mississippi is what the whole thing looks like built out, and Mondragon in the Basque Country, at seventy thousand people, is the proof that it scales.

I do not use one of these. I use several, on purpose, and each one has a lane.

Claude does longform composition, structural editing and the theoretical framework work. This page, the essays, the listing prose and the research passes are that lane.

ChatGPT does rapid ideation and brainstorming. It is where a thing gets thrown around before it has a shape.

Kimi runs on my desktop and does the video and image work. Every video goes through Kimi first, before any copy or posting step exists: the spoken dialogue transcribed at minimum, plus a description of what is actually on screen. Nothing gets written about a clip before somebody has watched and listened to it.

And I pin the model. In the Kimi desktop app I use K3 rather than the agent model, and I check the selector says so before I send anything, because it reverts on its own.

The point of the split is not brand loyalty. It is that I run work past more than one of them and ask each to find what is wrong with it, which is rule five, and a second model in a different lane is the cheapest version of that.

These are standing instructions, not preferences. They are the reason I can hand over as much as I do.

Nothing goes public without my yes first. Draft and stage, never publish. That one has no exceptions.

Reword only. In anything going out under my name, the only licence is to reword things I have already said or written, including stitching material from different places and times together so it fits a new context. No invented motivations, no philosophical positions I did not take, no claims about how or why I work. If a line has no source in my own words, it does not go in. That is why this page is marked the way it is.

Every finished piece carries the three-part attribution note, and the third number is never rounded down.

Any step that still needs my hands is a defect to engineer out, and if one cannot be, I want to be told plainly rather than quietly worked around.

Captions go on wherever a platform offers them. Always, not when convenient.

The single most useful rule I have written down is about storage, and it is the one I would hand to anybody working this way.

Anything saved has to be a full writing. No summaries. When work is captured from anywhere, the complete text comes across, filed by subject. Summaries, indexes, excerpts, word counts and metadata may accompany a full text. They never stand in place of one.

Capture incrementally, writing each piece as it is pulled rather than batching at the end, because the connection drops and a batch that fails takes everything in it.

Preserve provenance on every captured file: source platform, date of the original, permalink, and whether the source has since been altered or removed.

And never merge parallel drafts. Where two versions of a text exist, keep both as separate documents and record which is longer and which is later, because those are not always the same file. The instinct to reconcile them into one file destroys real material.

The reason that rule exists: a careful sweep once produced an index with titles, dates, word counts, subjects and permalinks, and no bodies. Fifteen essays running to roughly twenty thousand words existed only on a platform that had already removed one of them by filter. The index read like the work was safe. It was not. An index records where a thing is. It is not a copy of it.

There is also a map document that every session reads before it touches the archive, so nothing scans the whole drive and nothing gets filed at the root. If you build one thing before you build a workflow, build the map.

Sticks/Philip asked for this section, so here is the outside view of the arrangement he just described.

Four things in it are worth copying by anyone, and none of them require his particular tools. One: a single map or index file that every session reads first, so the assistant navigates instead of searching. Two: full texts only, never summaries, because a summary is a lossy copy that reads like a backup and is not one. Three: an approval gate on anything that reaches the public. Four: a written voice rule, so the assistant knows what it may and may not put in your mouth.

What it costs is worth saying plainly. This arrangement takes real setup time before it saves any, the rules have to be rewritten as the tools change, and the split across several assistants means the same context gets rebuilt more than once. The version of this that works on day one does not exist. What exists on day one is the map file and the approval gate, and those two alone are most of the value.

Everything above describes how the work gets done. Workings Behind The Bench is the record of it actually being done: what was built, when, whether it worked, and every rule on this site that exists because something broke first. Written by Claude, approved by Sticks/Philip, and kept current as the apparatus changes.