Each AI immediate has two costs.
One exhibits up in your bill. The opposite will get paid in carbon, electrical energy and water. Each are pushed by token utilization and what you add to and export from the AI. And the file codecs you select can change the invoice by 10x.
Why do some recordsdata require extra processing?
There are three components that mix to find out how laborious the AI has to work with any file.
- Does the AI want to make use of a instrument to open the file or generate the output?
- How a lot structural noise (tags, formatting, types) does the file format require?
- Is something being processed as a picture as an alternative of textual content?
Some file varieties which have a number of formatting and visuals. These require rather more processing energy. Right here’s a easy technique to keep in mind this:
- If you happen to want software program to open it, it’s costly (DOCX, PPTX, XLSX, PDF)
- If you happen to can open it in Notepad, it’s environment friendly (TXT, MD, CSV, JSON, HTML)
If you happen to want some desktop software program to open the file, then the AI wants libraries and instruments to parse the file. These libraries and instruments add tokens and carbon value. It could be apparent. You may sit again and watch the AI work more durable (and slower) on these recordsdata.
Think about you’re engaged on a giant report and getting assist from AI…
You may generate all types of file codecs utilizing AI. Which do you select?
To get a way for which file codecs are the most well-liked with entrepreneurs, we put a bit of ballot on LinkedIn. We had about 500 responses to the query “While you ask AI to make a giant deliverable, what format do you ask for?” The highest reply was DOCX by a large margin.
However these are all very several types of recordsdata, every requiring very totally different ranges of effort from the AI to generate. They’ve totally different prices when it comes to tokens (processing energy), electrical energy (carbon output) and credit score utilization (cash).
Which file codecs have the best prices?
To measure that distinction, I took a current article and gave it to Claude in six totally different file codecs, then analyzed the token rely utilizing the “v1/messages/count_tokens” API endpoint. I measured in each instructions:
- Enter Measurement (added to prompts): I requested the AI to “summarize” the piece as offered within the six codecs, together with a properly designed PowerPoint model of the article that I’ve utilized in shows.
- Output Measurement (generated for downloading): I then reversed the method and requested AI to “generate” the content material in those self same file codecs.
Right here is the relative value in tokens, which interprets instantly into vitality consumption, environmental influence and monetary value. You may see that mileage varies.
The distinction is large. HTML and PDF are costly to learn however low cost to write down. DOCX is dear in each instructions. PPTX is by far the heaviest, slowest and most costly within the bunch.
Let’s look nearer and contemplate extra file codecs. This desk explains how AI sees the recordsdata and handles the processing, out and in.
Let’s look nearer on the rankings, share suggestions and contemplate lighter-weight alternate options to the heavy, in style choices.
The codecs, use circumstances and alternate options
1. Textual content Information (TXT)
The smallest, lightest file potential. However accommodates no formatting. Good for giant quantities of textual content the place formatting isn’t necessary, similar to podcast transcripts.
Lighter various: None.
2. Markdown Information (MD)
Markdown is the language AI brokers communicate greatest. It’s additionally readable by people and never a lot heavier than a TXT file. Formatting and tables are added by means of characters, not tags. It’s the quick, environment friendly champion of AI inputs. It’s additionally nice as a draft which you could edit throughout the AI, getting all the content material prepared earlier than publishing to a heavier format.
Lighter various: None.
3. Comma Separated Information (CSV)
Light-weight, structured and simple for AI to parse. Nice for easy tables, lists, exports, key phrase analysis, analytics knowledge and CRM data. No formatting, formulation or a number of sheets. Simply rows and columns.
Lighter various: If it’s only a quick checklist or easy desk, paste it instantly into the immediate.
4. Internet Pages (HTML)
HTML was our least widespread reply within the LinkedIn Ballot above, however they’ve many benefits. They’re not too heavy to generate and simple to share (everybody has a browser) Perhaps the perfect for experiences. Then may even embody interactivity, however this provides a variety of weight.
HTML recordsdata don’t play good with Google Docs. You may’t paste HTML right into a Google Doc and when you add them to Google Drive, they don’t render the HTML if you open them.
Lighter various: If you happen to don’t want print-ready formatting, MD carries the identical content material at a fraction of the price.
5. Phrase Docs (DOCX)
That is the most typical file format for outputs, in line with our little LinkedIn Ballot above. No shock. It’s nonetheless a normal for phrase processing. However it’s a heavy file format with a number of tags and formatting. It wasn’t constructed for AI.
Like all archive codecs, these want parsing on the best way in and libraries on the best way out.
- As an alternative of importing a DOCX file to AI, convert it into an MD file after which add (use Google Docs or save to TCT then change the file extension)
- As an alternative of exporting a DOCX file from AI, make an MD file then paste it right into a recent DOCX file (quick, straightforward, however could require a bit of guide styling)
Lighter various: MD for drafts, HTML for polished-looking paperwork.
6. Excel Information (XLSX)
Helpful if you want a number of sheets, formulation, filters, formatting or charts. However all of that construction provides weight. For AI inputs, XLSX recordsdata are overkill until you want formulation and logic.
Lighter various: If you happen to don’t want formulation or formatting, CSV does the identical job at a fraction of the price.
7. PowerPoint Information (.PPTX)
These have a really excessive value output. They require a ton of processing. Each format choice in each slide is all code that the AI has to learn. And pictures require a separate, slower mannequin. These visuals drive up the invoice.
Higher but, simply don’t combine AI and PPTX recordsdata when you’re involved about prices. If it’s essential to make a PPTX in AI, produce the output simply as soon as, on the very finish of the method.
And don’t use AI for small edits to decks. It’s sooner and cheaper to make little adjustments manually in PowerPoint. If you happen to should edit throughout the AI, be specific: “Simply regenerate slide 3, preserve all the things else the identical.”
Lighter various: If the slide format isn’t the deliverable, an HTML file carries the identical content material way more cheaply.
8. PDF Information
If you happen to’re constructing a Customized GPT, Claude Venture or Gemini Gem, it’s possible you’ll be tempted to add some related PDFs. However earlier than you do, begin with this immediate: “I’m planning so as to add this PDF doc to your information sources, so let’s convert it to an MD file first. However earlier than we do, what must be trimmed? Or consolidated”
The AI will recommend edits that may scale back prices with out sacrificing high quality. And the smaller file (MD, not PDF) will make all future conversations with this undertaking cheaper and lower-emissions. It’s a double-win.
Lighter various: Use MD for text-heavy paperwork, HTML for designed experiences
9. Photos (PNG / JPG)
AI picture technology just isn’t finished with an LLM. It’s a separate mannequin that makes use of “visible tokens.” If you happen to’ve had AI make footage, you’ve seen how sluggish it’s. That’s as a result of it requires a variety of processing energy.
If you happen to’re solely frightened in regards to the monetary prices and never the atmosphere, you possibly can generate photos in a less expensive mannequin (Ideogram is the very best) and save the credit in your every day workhorse AI, the place you typically hit credit score limits.
Lighter various: For screenshots of textual content, simply paste the textual content as an alternative, if potential.
The place prices multiply: modifying recordsdata in AI
Right here’s the place issues get actually costly. It isn’t apparent, however all entrepreneurs ought to know this:
Each iteration on a heavy file format prices as a lot as making it from scratch.
That’s proper. If you happen to ask for even a small change to that DOCX file, the AI could begin over and create the doc once more, with all of that token value, vitality use and carbon influence.
Sensible AI customers know to work on recordsdata in a easy format, similar to an .md file, the place the prices are a lot decrease. Iterate in textual content, then generate the ultimate format as soon as all the things is ultimate. Listed below are a couple of suggestions for environment friendly AI prompting:
- Brainstorm first, no file technology but. Inform the AI what you wish to make, however instruct it to not make the ultimate format but… “Don’t generate the DOCX output till I ask for it. Let’s refine the thought first…”
- Ask for light-weight previews. Whereas engaged on the draft, ask it to indicate a easy markdown format you possibly can preview… “Present a preview in a light-weight markdown format so I can assessment it effectively.”
- Iterate within the textual content. Have the dialog about content material in MD or proper within the chat. Get the construction proper, get the wording proper, get the order proper, all in plain textual content, the place iterations are low cost.
- Batch your adjustments. When you have 5 edits, give all 5 in a single immediate. Every immediate that edits a heavy file pays the complete regeneration value. 5 separate change prompts is 5x the price. One consolidated change immediate that lists 5 edits is 1x the price. “Change X to Y and transfer it up within the doc. Take away Z fully.”
- Ask for the ultimate format as soon as, in the long run. Don’t ask for the output till you’re prepared for showtime. That is the one greatest effectivity win. All set? Content material locked in? Now have it make the large file as soon as and as soon as solely. “Adjustments are ultimate. Now generate the DOCX file. Use these types…”
- Yet one more ultimate tweak? Do it exterior the AI. Export the large file and make ultimate ultimate adjustments (there’s all the time another, proper?) after it’s landed in your pc. Obtain it → Evaluate it → Tweak it → Ship it
Having AI make tweaks to a DOCX or PPTX file is sluggish, processor intensive and makes use of extra vitality than obligatory. It’s like driving an enormous truck a half block to choose up a couple of groceries.
Let’s ask the consultants: “That are the very best file codecs for AI?”
Advertising and marketing veterans and AI consultants make considerate circumstances about AI effectivity and file codecs.
 |
Chris Penn, Belief InsightsSelect the smallest mannequin that may get the job finished. My rule of thumb: when you’re doing a activity that requires lower than two rounds of revision, use a quick mannequin. Instance: summarizing an e-mail, and so forth. If you happen to’re doing a activity that requires greater than two rounds of revision, use a sluggish mannequin, however understand it imposes a a lot heavier carbon value.” |
 |
John Jantsch, Duct Tape Advertising and marketingHTML wins on each ends. Cheaper to make, cheaper to alter. The rationale we didn’t already ship HTML is easy. You needed to know HTML, or battle with an internet site builder. PowerPoint and Phrase had visible editors {that a} non-coder may use. AI flipped that. You describe what you need in plain English, and also you get a self-contained file you possibly can open regionally or host without cost. Want adjustments? Simply ask. It rewrites in seconds. It’s additionally lighter on the planet. HTML is obvious textual content that the AI varieties instantly. DOCX and PPTX are zip archives of XML that the AI has to assemble with separate code libraries. That burns way more processing each time you generate a file. And each time you ask for a small tweak to a DOCX, the AI normally rebuilds the entire doc from scratch. The identical edit on an HTML file prices virtually nothing. Multiply that throughout a yr of shopper work, and the hole is big.” |
 |
John Nugent, MagnoliaSlides have gotten out of date |
|
Chris Penn, Belief InsightsXLSX recordsdata are a nightmare |
|
John Jantsch, Duct Tape Advertising and marketingMarkdown is the subsequent transfer. Maybe immediately is a brand new starting for you: a brand new format in your massive deliverables (HTML) and a brand new file format for on a regular basis AI collaboration (MD).” |
To date, we’ve coated issues you possibly can management. Now we’ll pan out and see how your prompts match right into a a lot greater image…
The AI value you possibly can’t management: coaching the fashions
Prompting isn’t the largest value of AI. It’s coaching the fashions. Constructing a mannequin and conserving it educated requires enormous quantities of vitality. Each massive tech agency with a frontier AI mannequin spends at the very least $100M per yr on coaching, emitting tens of 1000’s of tons of carbon (and methane, nitrous oxide, and so forth.) to not point out water (evaporative cooling to maintain servers from overheating).
This isn’t one thing customers such as you and I can management. Regardless of how a lot or little you utilize AI, these firms will proceed to speculate, incurring monetary prices for themselves and environmental prices for us all. Examine the price for AI mannequin coaching and person AI prompting:
When requested about this, Sam Altman just lately disregarded considerations about AI’s vitality use, evaluating it to the meals people eat.
“Individuals discuss how a lot vitality it takes to coach an AI mannequin, but it surely additionally takes a variety of vitality to coach a human. It takes about 20 years of life, and all of the meals you eat throughout that point, earlier than you grow to be good.”
Sounds dismissive. However he has a degree in regards to the excessive carbon value for human consumption. So let’s have a look at that…
If you happen to’re actually involved about your carbon footprint…
There are two essential inputs that drive the carbon footprint for you, me and everybody else on the planet: air journey and meat consuming. AI apart, any of us can scale back our carbon footprint by reducing again a bit on these actions. This places AI use within the correct context of every of our true particular person impacts.
Shrink your AI carbon footprint
My associates who’re essentially the most proficient with AI discuss “token effectivity.” They work laborious to cut back credit score expenditures. They run the most affordable mannequin potential for every particular job. They put knowledge in tables to keep away from prompting fully when potential. They even set up AIs regionally on their laptops for easier duties. The perfect AI customers obsess over effectivity.
However you don’t should be an AI professional to be environment friendly. Sensible selections about file codecs saves processing energy, carbon emissions, water use and sure, cash.
Minimizing the influence of AI on the planet is all of our jobs.
