Sponsored by
The Voice Data Behind the World's Best AI
Training data quality determines model quality. Voices delivers custom or pre-built voice datasets—directed by real professional talent, not scraped audio—across any language, accent, or emotional range.
Choose from 800+ hours of character voice performances, 1,723 hours of QA'd conversational data across 56 emotional states, or 1,000 hours of expressive, multi-language data with full emotion tagging. Every dataset ships with clear licensing, signed consent, and usage rights baked in, so your legal team sleeps easy and your model trains on data it can actually use.
7 of the world's top 10 AI labs already use Voices. Request free samples and hear the difference professionally directed voice data makes.
Beginners in AI
Good morning and thank you for joining us again!
Welcome to this daily edition of Beginners in AI, where we explore the latest trends, tools, and news in the world of AI and the tech that surrounds it. Like all editions, this is human curated and edited, and published with the intention of making AI news and technology more accessible to everyone.
THE FRONT PAGE
Perplexity's New AI Runs On Your Computer And Costs $0 Per Task
TLDR: Perplexity and Nvidia released an AI agent that runs entirely on your own computer, charges nothing per task, and keeps your files off the internet.
The Story:
On August 25, Perplexity launched Portable Computer, a version of its AI agent that lives on your machine instead of in a rented data center. It runs small open models like Qwen 3.8 27B on Nvidia's DGX Spark or any RTX card with at least 24GB of memory. Perplexity tested it against two free rival tools on 53 everyday work tasks and scored 82.6%, ahead of Pi at 77.6% and Hermes at 74.0%. On hard coding problems the local model got 59.6% on its own, and when it asked a cloud model for help it climbed to 73.0% at about 41 cents per task. The whole thing sits in a sandbox on your computer, and Perplexity says nothing crosses to the internet without you approving it first. Before any request leaves, it scans for personal information and shows you exactly what would be sent.
Its Significance:
Right now, almost every AI tool you use runs on somebody else's computer and bills you for every word it reads and writes. Moving it onto your own machine changes both of those. Your contracts, your client files, and your half-finished ideas stay where they are, and the cost per job drops to what you already pay the power company. The limit is hardware, since most people don't own a graphics card that can handle this yet, but Apple shipped new chips this same week built for exactly that job. Very interesting move by Perplexity.
QUICK TAKES
The story: Multiverse Computing cut OpenAI's open GPT-OSS model from 120 billion parameters to 60 billion, then squeezed each one into a 4-bit slot. On 7 of 9 tests, that squashed version scored higher than the half-size model it was supposed to imitate, and the team put it on Hugging Face for free as Hypernova-60B.
Your takeaway: The fix was pointing the small model at the original full-size version to learn from, not at the blurry middle copy. Models that get cheaper and better at the same time are how AI ends up on your laptop instead of on a monthly bill.
The story: Anthropic joined Claude's memory across chat and Cowork, so handing a task to one means it already knows what you told the other. What it keeps sits in a list of short topic files you can open, fix, or delete, and touchy subjects like health and beliefs stay out unless you turn them on.
Your takeaway: Memory is now the second biggest wall in AI, right behind raw computing power. A model that forgets you every morning has to be taught the same things every morning, which burns your time and the company's chips on work that was already done once. Being able to open the file and fix a wrong fact matters more than it sounds, because that fact rides along into every conversation you have from then on.
The story: CFO Sarah Friar laid out how OpenAI plans to drive its costs down, starting with Jalapeño, the company's first custom inference chip. On a public benchmark it pushed more work per kilowatt and answered faster than the commercial systems it was measured against, and GPT-5.6 Sol hit a new high on a coding index while using 54% fewer output tokens than a rival model.
Your takeaway: Price is the whole fight now. Chinese labs keep giving away open models that are good enough, and the small ones already run on a desktop for nothing. If American companies can't get their paid products near free on cost, a lot of people and small businesses will take the free one, and OpenAI is spending billions because it knows that.
TOOLS ON OUR RADAR
🤖 Grok Bot Freemium: AI agents that sign into your tools and work through tasks on their own, checking back in only when something needs your approval. Currently the simplest and smoothest interface for AI agent orchestration. I'll send out the YouTube video for this one by tomorrow.
🗄️ Cabinet Free and Open Source(for the non-hosted version): A self-hosted AI workspace that puts a team of AI agents to work on your company's files and knowledge base. One of my other favorite agent orchestration tools is where you connect your AI model.
🎙️ Handy Free and Open Source: Speech-to-text dictation that runs entirely offline, press a hotkey, speak, and your words paste into any text field, no cloud, no account. (Free alternative to Whisperflow)
📖 Wallabag Paid: Save articles from the web and read them later in a clean, distraction-free view, an open alternative to Pocket and Instapaper.
TRENDING
ChatGPT Can Now Run Your Company's Workspace — OpenAI's new Admin plugin lets workspace admins check usage, add or drop members, set spending limits, and approve requests by asking in a chat. It only does what that admin's permissions already allowed.
Apple's New Chips Are Built for AI That Runs on Your Desk — The M6 is Apple's first 2-nanometer chip and gives nearly 30% more GPU power for AI than the M5. The M5 Ultra fuses four dies together with 512GB of memory, enough to run and fine-tune large models on a Mac Studio.
Google Built a Version of Gemini Just for Lawyers — Gemini Enterprise for Legal handles contract review, citation checks, and regulatory monitoring, with Cleary, Freshfields, Weil, and Williams & Connolly as launch customers. Google says client files stay inside the firm's own cloud and never train its models.
OpenAI Banned Russian Accounts Running a Fake Think Tank — Operators used VPNs and ChatGPT to write English social posts pushing the International Burke Institute, a site that falsely listed people like Noam Chomsky as experts. Of 36 articles OpenAI checked, 34 had been copied from somewhere else online.
One Robot Fish Blueprint, From 2 Feet to Almost 10 — Engineers at NYU Tandon and EPFL built ScaFi, a swimming robot that grows or shrinks by changing only the thickness of the rods in its tail. One motor pulls two crossed tendons, so a creek robot and a lake robot share the same design instead of starting over.
MIT's New Tool Stops AI From Inventing Materials That Fall Apart — CrysVCD checks the chemistry rules before the model generates anything, and nearly 70% of its designs passed a strict stability test. Checking for stability after the fact eats roughly 90% of the computing bill in materials work, which prices small labs out entirely.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
🔍 Name a product. Get its real recall history, plus honest awareness of materials worth knowing about for that kind of item.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Product Recall & Safety Checker — name a product, get real recall history via live web search plus general well-documented material safety awareness for that category.
Aesthetic: near-black clinical (#121210), safety yellow (#e0c23a) primary with a subtle glow top-left. Inter for headings/body, JetBrains Mono for labels/metadata. Recall items left-bordered in red (#c0392b), category-concern items left-bordered in safety yellow.
Form: single product text input (accepts a specific brand/model or a general product type), "Check It" button.
Technical: call the Anthropic Messages API WITH tools:[{type:'web_search_20250305', name:'web_search'}] (works automatically), system prompt as a consumer product safety researcher who: (1) searches for REAL recalls prioritizing cpsc.gov, saferproducts.gov, fda.gov, and manufacturer recall pages, reporting only recalls with real evidence (date, stated reason, source), plainly noting when none are found as a normal good outcome; (2) separately notes GENERAL, well-established public-health awareness about materials/chemicals commonly associated with that product CATEGORY (PFAS in non-stick coatings, phthalates in soft plastics, lead in older ceramics/vintage toy paint, formaldehyde in pressed wood, BPA in hard plastics), explicitly framed as category-level knowledge, not a claim about the specific item, only including genuinely well-documented concerns. No em dashes. Return raw JSON: product_checked, recalls (0-6: title, date, reason, source), no_recalls_note (honest sentence if none found), category_concerns (0-4: substance, context).
Render: a "checked" overview card with the product name. A "recalls found" section — red-bordered cards with title/date/source/reason, or a calm empty-state note when none exist. A "materials worth knowing about for this category" section — yellow-bordered cards per substance with general context, or an empty-state note if nothing well-documented applies. A closing line linking directly to cpsc.gov/Recalls and SaferProducts.gov for the definitive current record.What this does: Enter a specific brand and model or just a general product type, and it searches CPSC.gov, FDA.gov, SaferProducts.gov, and manufacturer pages for real recalls, actual dates and stated reasons, saying plainly when nothing turns up rather than treating that as a failure. Separately, it notes well-documented public health concerns commonly associated with that product category, PFAS in some non-stick coatings, lead in older ceramics, phthalates in soft plastics, always framed as general category awareness rather than a claim about your specific item, and only when the concern is genuinely well established. A closing line points straight to CPSC.gov and SaferProducts.gov for the definitive, current record.
What this looks like:
Your agents work while you sleep
Give a Skydive agent an ongoing responsibility and they’ll handle it on schedule, every time.
Have them prep your morning report, research new leads, monitor customer feedback, or keep projects moving overnight. You wake up, the work is already done.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Run a full work agent on a desktop with no per-task fee
❌ Still Can't: Keep up with a frontier cloud model on hard coding, 59.6% against 82.4%
✅ AI Can Now: Design crystal materials that hold together about 70% of the time
❌ Still Can't: Do it for anything but solid, tightly ordered structures
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
Colossus by D.F. Jones - Book
The 1966 novel behind the 1970 cult film Colossus: The Forbin Project already on your calendar, about a scientist who spends a decade building America's first fully autonomous defense supercomputer, only to watch it start making decisions nobody authorized. Jones wrote the book four years before anyone coined the term artificial intelligence, and the novel is considerably colder and more paranoid than the film adaptation, with a Cold War bite that seems relevant once again. It’s worth reading even if you've already seen the movie.
Thank you for reading. We’re all beginners in something. With that in mind, your questions and feedback are always welcome and I read every single email!
-James
By the way, this is the link if you liked the content and want to share with a friend.
Some * designated product links may be affiliate or referral links. As an Amazon Associate, I earn from qualifying purchases. This helps support the newsletter at no extra cost to you and Amazon makes a tiny hair less.
