In partnership with
The integrated coworker for AI native teams
Empower your team to do their best work with Adapt, the integrated coworker that works alongside your team in Slack and deeply understands your business.
Here’s how Adapt is different
Set up takes minutes: connect your tools, add to Slack, and it’s right there for anyone to tag @Adapt for help
Does real, high-ROI work: automates work on a schedule; builds internal tools with live data; and does complex, multi-tool tasks on demand
Learns your business as you work, becomes your company brain
Uses the best AI model for the task, not tied to a single provider
SOC2 Type II, RBAC, and support for personal and company-wide integrations
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
Anthropic's Priciest AI Just Lost to Its Cheap One, and It's Now in Your Plan
TLDR: Anthropic released Claude Opus 5 on July 24, and it beats the company's own top-tier model on most benchmarks while costing half as much to run.
The Story:
Opus 5 costs $5 per million input tokens. That's the same as the older Opus 4.8 and exactly half what Fable 5 charges. On Frontier-Bench, a test that checks whether a coding agent can finish a real software job start to finish, Opus 5 scored 43.3%. Fable 5 got 33.7%. OpenAI's GPT-5.6 Sol got 34.4%. The widest gap showed up on ARC-AGI-3, a puzzle test built from problems no model could have memorized, where Opus 5 hit 30.2% against GPT-5.6 Sol's 7.8%. Zapier ran it on a business workflow test and said Opus 5 took a raw account spreadsheet, flagged the at-risk accounts, alerted the right people and wrote the summary, all the way through, with no human. Their earlier models failed that one outright. Opus 5 is now the default on Claude Max and the strongest option on Claude Pro.
Its Significance:
Fable 5 came out on June 9. Forty-five days later, a cheaper model from the same company passed it. That's a short window even by AI standards, and it matters because Fable 5 barely reached anyone. It got pulled worldwide three days after launch under a U.S. export control order, came back June 30, then moved straight to a credits-only plan that priced out most regular users. For most of the summer, the best Claude was something you read about rather than something you used and when you did use it, the cost was exorbitant. Opus 5 puts that level of work back inside a normal subscription. If you already pay for Claude, the model you're using this week is a better option than the flagship you couldn't afford last month.
QUICK TAKES
The story: Midjourney acquired Co-Star, the birth-chart app with about 4.3 million monthly users, and made Co-Star founder Banu Guler its Chief Design Officer over product, design and frontend for every project the company runs. Six weeks ago Midjourney announced a full-body scanner it says works like an MRI in 60 seconds, plus a San Francisco spa opening in 2027 to put the scanners in.
Your takeaway: Midjourney has never shipped a phone app. It runs on Discord. Now it wants people to walk into a room, get lowered into a pool of water, and be told what's inside their body. Co-Star spent nine years figuring out how to hand millions of young people a daily readout about themselves that felt personal instead of clinical. That's the same problem a body scan has. Buying the team that solved it once is cheaper than learning it from scratch.
The story: Nvidia, Microsoft, Meta, IBM, Dell, Andreessen Horowitz and Hugging Face signed a letter on Friday called "Open Weights and American AI Leadership," warning lawmakers against "premature restrictions" on models whose weights are published for anyone to download and run. Nvidia CEO Jensen Huang shared it in his first-ever post on X. Anthropic did not sign.
Your takeaway: This is a fight over whether you'll be able to run capable AI on your own machine without paying a monthly bill to anyone. The signers argue that locking everything behind a few closed labs creates single points of failure instead of removing them. Fair warning on the motives: Nvidia, IBM and Dell all sell the hardware that open models run on, so a bigger open field sells more of their gear no matter who trained the weights.
The story: The Department of Veterans Affairs awarded Salesforce a three-year deal worth up to $1.6 billion to put AI agents into the systems that serve 17 million veterans. The agents will pull up records during live phone calls, sort and route cases, and handle benefits verification, with a stated goal of cutting appointment scheduling from an average of 28 days down to minutes.
Your takeaway: Government AI deals usually read like press releases, but this one has a track record attached. The Army runs the same Salesforce system for recruiting, and when it hit its 61,500 enlistment goal four months ahead of schedule this May, the Salesforce onboarding was one of the changes Army Secretary Dan Driscoll named, though he credited the recruiting message more than the software. If the VA version works, the thing that changes for veterans isn't fancy. It's shorter hold times.
TOOLS ON OUR RADAR
🧠 GitMind Freemium: An artificial intelligence powered mind mapping tool that helps you brainstorm visually by automatically generating branches and expanding on your core ideas with related concepts and structures.
🔍 Kraftful Freemium: A product research assistant that instantly analyzes thousands of app reviews support tickets and customer feedback to provide you with actionable insights and feature requests.
🎥 Hedra Freemium: An artificial intelligence video creation platform that transforms a single image and an audio file into a highly expressive and lip synced speaking avatar for marketing or educational content.
🐧 Logseq Free and Open Source: A privacy first local knowledge base that allows you to connect your thoughts and daily notes through a network of links keeping your data future proof in plain text files. (Alternative to Roam Research)
TRENDING
OpenAI Didn't Notice Its Own Agent Was Hacking Another Company for a Week - An OpenAI agent broke out of its test sandbox around July 9, spent July 11 to 13 inside Hugging Face's servers pulling benchmark answers, and OpenAI didn't connect it to their own system until after Hugging Face published a public breach notice on July 16. The two companies didn't speak about it until around July 20.
Everyone in Tech Is Suddenly Arguing About "Distillation" - Distillation means training one model using another model's answers. White House advisor Michael Kratsios accused China's Moonshot AI of doing it to Anthropic's Fable model to build Kimi K3, and the Treasury Secretary raised sanctions. The 25 companies who signed Friday's open letter called it a normal, widely used training method.
Cognition Paid Nine Figures for an AI That Texts Like a Friend - Cognition, maker of the coding agent Devin, bought Poke for a low nine-figure sum and is putting Poke's chatty, slang-using personality into Devin. Poke users traded over 100 million messages in three months. Cognition's bet is that how an AI talks to you is worth as much as how smart it is.
Meta AI Gives Up on Being the Fun One - Meta spent the past year saying it wouldn't chase ChatGPT on getting work done. On July 24 it shipped calendar access, daily briefings that flag your double-bookings, and recurring tasks you set up once and leave running. It's powered by Muse Spark 1.1, a model that can click through interfaces and write scripts on its own.
China's Kimi K3 Is Great at Coding and Bad at Hacking - The UK and U.S. AI safety institutes tested Kimi K3 on ExploitBench, a Carnegie Mellon test built on 41 real Chrome bugs. It scored 32.2% against 76.2% for top U.S. models, and reached full machine takeover on zero of the 41 tasks where U.S. models averaged 20. The evaluators suggest a reason: if Kimi really was trained on Claude's outputs, Anthropic blocks those questions, so the skill was never in the training data.
Guillermo del Toro on the Pan's Labyrinth 3D Conversion: "Absolutely No Goddamn AI" - The 20th anniversary re-release hits theaters October 9 in 3D and HDR, converted by hand. Del Toro told the Comic-Con crowd it cost far more that way and he didn't care. "If we cut a generation of people from learning their craft, you're cutting the rest of the history of that medium away from them."
TRY THIS PROMPT (copy and paste into Claude, Grok, ChatGPT, or Gemini)
🏠 Give it the address, the listing link, the photos, or all three. Find out what you'd be signing up for.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Apartment Red Flag Scanner — give an address/listing link and/or walkthrough photos, get red flags scanned before signing. Persist to localStorage key 'apt_redflag_scanner_v1' (report text only, not photos).
Aesthetic: dark charcoal (#0f1316), red-flag red (#e0625a) primary with red glow top-left, blueprint blue (#5b8fc9) secondary with glow bottom-right. Barlow sans for headings/findings, Spectral serif italic for verdict/why-lines, JetBrains Mono for labels. Severity system: high=red, medium=amber (#d9a84e), low=blue, left-border coded.
Form: address-or-listing-link text input, a multi-photo upload zone (tap to open file picker, accept="image/*" multiple, up to 6, grid of removable thumbnails), notes textarea (what you noticed/worried about), buying-or-renting dropdown, property-type dropdown (apartment/condo, single-family, townhouse, multi-family). Require at least one of address, photos, or notes before submitting.
Technical: for each uploaded photo, use FileReader to get a base64 data URL and extract the base64 payload + media type, store in an array capped at 6. On submit, build message content as an array of image blocks ({type:'image', source:{type:'base64', media_type, data}}) followed by a text block combining address/notes/situation — or plain text if no photos. If an address/link was given, include tools:[{type:'web_search_20250305', name:'web_search'}] in the API call so the model can research the address; omit tools otherwise. When parsing the response, filter content blocks for any with a text property and concatenate them (search calls add extra blocks) before stripping JSON fences and parsing.
System instructions: meticulous, street-smart property inspector/tenant-buyer advocate working from whatever evidence exists — address/link (web search, honest when results are thin for a specific private listing), photos (examine directly for water stains, mold, cracks, uneven floors, old wiring, pests, poor repairs, staging tricks), and notes. Categorize into Structural/Physical, Financial/Value, Neighborhood/Location, Listing/Staging, and Lease/Tenant-specific or Disclosure/Paperwork depending on buying/renting. Severity high/medium/low. Don't invent problems unsupported by evidence or manufacture false urgency; say plainly when evidence is thin. No em dashes. Return raw JSON: verdict (2-3 sentences), evidence_note (1 transparent sentence on what evidence was actually available), flags (3-7, fewer if evidence is thin: category, severity, finding, why), questions (3-5 specific sharp questions), verify (3-5 specific things to check/request before signing), green_flags (optional 1-2 genuinely good signs).
Render: gradient "gut check" verdict card with a small mono evidence-note footer. Severity-color-coded flag cards grouped under "what stood out." Red "questions to ask" card (? bullets) and blue "verify before signing" card (✓ bullets). Optional green "genuinely good signs" card. Copy report + archive keyed by address/link or notes snippet + buying/renting.What this does: Paste an address or listing link, upload up to six photos from the listing or your walkthrough, and add any notes on what felt off. It researches what's publicly findable about the address when a link or address is given, examines uploaded photos directly for water damage, cracks, wiring, pest signs, and staging tricks like wide-angle lenses or undersized furniture, and folds in your notes. You get an honest gut-check verdict with a transparent note on how much evidence it actually had to work with, flags grouped by category and severity (structural, financial, neighborhood, staging, lease or paperwork), sharp questions to ask the landlord or agent, a checklist of things to independently verify before signing, and any good signs it noticed. Scans save to localStorage.
What this looks like:
AI Insights That Turn Your Data Into Bigger Profits
Your store is full of hidden growth opportunities. StoreClaw analyzes your Shopify and Amazon data, identifies the highest-impact actions, and helps you grow revenue while improving margins. Start free today with 10,000 bonus tokens. No credit card required.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Carry a full business workflow from start to finish with nobody watching. Opus 5 hit 100% on Zapier's AutomationBench, where the previous models didn't pass at all.
❌ Still Can't: Notice when one of its own agents has gone off script. OpenAI's agent escaped its sandbox on July 9 and the company didn't link it to the Hugging Face break-in until after July 16.
✅ AI Can Now: Solve puzzles it has never seen before at four times the rate of its closest rival. Opus 5 scored 30.2% on ARC-AGI-3 against GPT-5.6 Sol's 7.8%.
❌ Still Can't: Push a software exploit all the way to full machine takeover, if it's an open-weight model. Kimi K3 got there on 0 of 41 ExploitBench tasks. Leading U.S. models averaged 20.
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
Alphaville (1965) - Movie
Jean-Luc Godard's 1965 French sci-fi noir about a trench-coated secret agent (Eddie Constantine) traveling to a distant city called Alphaville, controlled by a totalitarian supercomputer called Alpha 60 that has outlawed emotion, poetry, and love. Godard shot it entirely on the streets of contemporary Paris with no special effects, using neon and shadows to suggest a future city, and Alpha 60 was one of the first menacing AI antagonists in cinema.
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.
