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
OpenAI Says ChatGPT-5.6 Beats Claude and Grok. Testers Say It's Complicated
TLDR: OpenAI shipped GPT-5.6, a three-model family named Sol, Terra, and Luna, and early tests show it leading on coding while costing less than its rivals.
The Story:
OpenAI put out GPT-5.6 on Thursday. It comes in three sizes: Sol, the flagship, Terra, the everyday one, and Luna, the cheap one. This is the first time OpenAI gave its models names instead of just numbers. The top model spent two weeks locked in a preview that the U.S. Department of Commerce limited to about 20 approved partners before opening it to everyone. On a test called Terminal-Bench 2.1, which measures how well a model handles command-line work, Sol scored 91.9% in its strongest mode. That beats Anthropic's Claude models and Google's Gemini 3.1 Pro. OpenAI leaned hard on cybersecurity too. On ExploitBench, a test of finding and using software flaws, Sol matched Anthropic's locked-down Mythos Preview while burning about a third of the tokens. Sol costs $5 and $30 per million input and output tokens, and Luna drops to $1 and $6. Developers who got in early were impressed. One tester summed it up: GPT-5.6 is a Porsche, and Anthropic's Fable is a warp drive.
Its Significance:
Five big labs shipped new models in seven days. OpenAI, xAI, and Meta all launched inside one week, which leaves Google's Gemini 3 as the oldest top model still standing. For you, more competition usually means better tools at lower prices. Sol costs less than Claude Fable 5 and handles more of the daily work for less money. The cheaper Luna beats Anthropic's Opus 4.8 on OpenAI's own coding test, so small teams and solo builders can get strong results without paying top dollar. The cybersecurity part cuts both ways. A model that's good at finding software flaws can help defenders patch holes faster, but that same skill worries security experts. OpenAI says Sol still doesn't cross its own "too dangerous" line.
QUICK TAKES
The story: The EU said Friday that Meta must change the addictive design on Facebook and Instagram or face a fine. The Commission pointed to infinite scroll, autoplay, push notifications, and personalized feeds that it says push people, especially kids, into compulsive use.
Your takeaway: If the ruling holds, Meta could pay up to 6% of its yearly global revenue, which is close to $12 billion. It could also mean the apps your family uses start to feel different, with fewer features built to keep you glued to the screen.
The story: Perplexity took a free Chinese model called GLM 5.2 and trained it to run its Computer agent, handing off the hard tasks to Claude Opus 4.8 only when needed. The result costs about a third of what Opus 4.8 costs while doing similar work.
Your takeaway: This shows a cheaper way to build AI: use a free open model for most jobs and call the pricey one only when you have to. For businesses, that points to lower AI bills. For free open models, it's proof they can now compete near the top.
The story: Sunrun is testing a plan to turn homes that have its solar panels and batteries into tiny AI data centers. It puts small compute nodes in customer homes, sells that computing power to companies, and pays homeowners for hosting the gear.
Your takeaway: If it works, people with rooftop solar could earn a new stream of income from the AI boom. It also skips the years of permits and construction big data centers need, since the power and the homes are already in place.
TOOLS ON OUR RADAR
🐧 BentoPDF Free and Open Source: A highly capable client side document toolkit that runs entirely in your web browser. It allows you to merge split compress and edit your files locally meaning your sensitive documents never touch an external cloud server. (Alternative to Adobe Acrobat)
🌐 Harpa AI Freemium: A hybrid artificial intelligence extension that lives entirely in your browser turning any website into an interactive data source so you can track prices summarize long articles and automate repetitive clicks.
🧠 Fabric Freemium: A spatial internet operating system that automatically organizes your chaotic bookmarks files and notes into visual collaborative spaces so you never lose track of important research.
✂️ Kome Freemium: A clever browser companion that condenses long YouTube videos news articles and lengthy email threads into digestible summaries giving you the core insights in a fraction of the time.
TRENDING
AI Tool Scours the Web for Job Openings, Preps Your Resume and Cover Letter - A developer built Autopilot-Jobhunt, a free tool that searches job posts while you sleep, scores them against your resume, and messages you the best matches on Telegram. It can draft a tailored resume and cover letter, but it won't apply for you; you review and send.
AI Agents Could Be Turned Into Botnets Through Hallucinations, Researchers Warn - Researchers from Tel Aviv University, Technion, and Intuit showed an attack called HalluSquatting, where hackers register the fake code links that AI coding tools tend to make up. When an AI agent later reaches for that made-up link, it can run the attacker's code. In some tests it fooled popular coding assistants close to 100% of the time.
Ethereum Foundation Turns AI Loose on ETH Network to Find Bugs Before Hackers Do - Ethereum's security team pointed swarms of AI agents at its own code and found a real flaw, now patched. The team said the hard part wasn't finding bugs but telling real ones from fakes, so a human still makes the final call.
AlphaEvolve Is Available for Everyone - Google opened AlphaEvolve, a Gemini-powered agent that writes and tests better algorithms, to all Google Cloud customers. Early users cut warehouse travel by 10.4% and sped up some model training, and Google used it to improve its own chips.
Bringing Ode Poetry to Life With MAI's Audio Models - Microsoft built Ode Poetry, an app that asks how you're feeling and reads you a poem in the voice of poetry expert William Sieghart. It runs on Microsoft's own MAI voice and transcription models, with the human poem and human reading kept at the center.
ChatGPT Is Now a Partner for Your Most Ambitious Work - OpenAI launched ChatGPT Work, an agent that pulls info from your apps and builds finished sheets, slides, docs, and web apps, sticking with big projects for hours. It runs on GPT-5.6 and can keep working on scheduled tasks while you're away from your desk.
TRY THIS PROMPT (copy and paste into Claude, ChatGPT, or Gemini)
📥 Paste the mess sitting in your inbox. Walk away with three sorted piles, and the reply-now drafts already written.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Inbox Zero Sprint — paste a rough inbox dump, get it triaged into reply-now, delegate, and archive, with drafted replies. Persist to localStorage key 'inbox_zero_sprint_v1'.
Aesthetic: dark teal-black (#0c1215), cyan (#3ec6e0) primary with top radial glow. DM Sans for headings/items, Source Serif 4 italic for overview/notes, JetBrains Mono for labels and the reply-draft blocks. Three bucket colors: coral (#e8735f) reply-now, violet (#a78bfa) delegate, slate-blue (#8ba8bc) archive.
Form: big inbox-dump textarea (one item per line: sender - subject - gist, placeholder shows a realistic mixed example), reply-tone dropdown (brief & friendly / warm / formal / direct), optional "who could you delegate to" text input.
System instructions: ruthless-but-kind inbox triage assistant. Sort EVERY item into exactly one of reply_now / delegate / archive. For reply_now, write an actual usable 2-line reply draft in the requested tone, specific to the content, no generic filler, no em dashes. For delegate, suggest a name from the given list (or a role if none given) plus a one-line handoff note. For archive, give a short reason it's safe to skip. Return raw JSON: overview (1-2 sentences), items[]{sender, subject, category, note (delegate/archive reason), reply (reply_now draft only), delegate_to (delegate only)}, start_with (1-2 sentences naming the first reply to send).
Render: gradient "the sweep" overview card with a 3-stat live-count row (color-coded). Three labeled columns (dot + title + subtitle) of left-border item cards. Reply Now cards show a mono reply-draft block with its own Copy button. Delegate cards show "→ **name**. note". Archive cards show the reason. Cyan "start with" closer card. Archive of past sprints keyed by a snippet of the dump.What this does: Dump a rough list of what's piled up, sender, subject, gist, one per line, messy is fine, plus your reply tone and who you could hand things off to. It sorts every single item into Reply Now, Delegate, or Archive. Reply Now items come with an actual usable 2-line draft in your chosen tone, ready to copy. Delegate items name who or what role should take it and how to phrase the handoff. Archive items get a one-line reason it's safe to skip. A quick overview names what it noticed about the inbox as a whole, live counts sit up top, and a closer names the single reply to knock out first. Saves each sprint to localStorage.
What this looks like:
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Score 91.9% on command-line coding tasks and match a locked-down security model while using a third of the tokens (GPT-5.6 Sol).
❌ Still Can't: Be trusted to tell a real software bug from a fake one on its own. Ethereum's team still needs humans to check every finding.
✅ AI Can Now: Run most agent tasks on a cheap open model and call a pricey frontier model only when a job actually needs it (Perplexity's tuned GLM 5.2).
❌ Still Can't: Be handed full permissions safely. The same hallucinations that cause wrong answers can be turned into an attack (HalluSquatting).
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
Feed by M.T. Anderson - Book
A 2002 YA novel set in a near-future where nearly every American has a brain implant called the "feed" that pipes advertising, entertainment, and social media directly into their thoughts, told from inside the head of a teenager named Titus who meets a girl who wasn't implanted until she was seven and thinks about the world completely differently. National Book Award finalist, weirdly funny, and poignant.
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.
