→ Noam Shazeer (at Google since 2000) left for OpenAI → Nobel Prize winner John Jumper left DeepMind for Anthropic → Now Gemini researchers Jonas Adler & Alexander Pritzel are heading to Anthropic too This isn't random. OpenAI and Anthropic are both preparing for IPOs, and pre IPO equity is a rare payday even for well paid Big Tech employees. Shazeer's exit came shortly after Google reassigned computing resources from his project to another team. When top researchers can't get GPU access for their own work... they leave. The real question is : does this signal a shift in where the most ambitious AI work is happening?

OMG! Did I just get burned. Easily my biggest ouch in the 18 months I've been using Make. I apologise in advance for the poor quality video, (a lag between my voice and screen) but hope you can hear my frustration and start to see my problem. In a nutshell, how do you get a scenario to process batches (of Airtable data) and then end cleanly when it's iterated through all the records. Alternatively, what AI Agent or AI model/LLM could I use (via OpenRouter) that would process 1000s of records). That would be the easiest solution.

Hi team. I can't believe I've wasted a whole day on this. I've worked with Claude Chat, Claude Cowork, Claude Chrome and ChatGPT and they've ALL FAILED to deliver a working solution; i.e a clean Google Doc output format path for a clean WordPress input format path. They had no restrictions on modules to use, or number of modules, but despite showing them the EXACTLY the output they needed from the input provided, I'm still no closer to success than when I started! FYI, here is the Google Doc which I've shared publically for now: https://docs.google.com/document/d/1pziCm9wIIvtKvzt10IRc_jk-5pl1pn0rd8RVrjsBOLM/edit?usp=sharing PS. Here is how Cowork described the clean input that is working inside the scenario: https://us2.make.com/93886/scenarios/4492233/edit?folder=63806 Note I don't know if the following IS the reason it works. All I know is that the single module scenario above with the clean "content" inside it, works. Claude trying to self-diagnose: "When you gave me the blueprint with the content already populated, it worked. Here's why: The "content" value is plain HTML — but with a very specific character encoding approach. Here's what's actually in there: What it IS: - Raw HTML tags: <h1>, </h1>, <p>, </p>, <h2>, <ol>, <li>, <table>, <tr>, <td>, <br>, <hr>, etc. - All on a single continuous line — zero newlines, zero line breaks anywhere - No escaped characters like \" — double quotes simply don't appear inside the content value at all What it uses for special characters: - Named HTML entities for typographic characters: &ldquo; &rdquo; &rsquo; &ndash; &nbsp; &pound; &reg; - Numeric entity &#39; for apostrophes (instead of a literal ') - &gt; for > where needed - &quot; for quote characters where needed The critical difference from what LLMs produce: When an LLM (or a Make module) processes text from a Google Doc, it almost always produces one or more of these problems:

It seems so simple! But after two days of trying to get this to work, I need to admit defeat. I've done nearly 40 experiments I've tried join, map formulas I've tried grouping I've tried iterating I've tried aggregating I've fought with Claude I've fought with ChatGPT I've read Make Community posts I've watched YouTube videos I've looked at other blueprints But still no clue as to how I can do this simple task! There must be a way. But I need a fresh set of eyes and a new way of thinking about the problem and the solution. How do I aggregate operations so that I can simply join a number of text strings together into one single string? Could I use a set variable and a replace formulawith a variable lifetime setting that does one operation at a time Could I get the output in JSON and parse it that way? Could I use a text parser? Could I use a tools 'increment' function to move through the operations one at a time? PS. I heard maybe a data store is the way to go. But that seems like an incredibly complicated solution.

I have a number of custom GPTs linked to my SEO tools. I don't use them all that often as most of my project work is executed with Claude. However, today I tried to use these tools but execution was continually thwarted. I eventually asked the question are all my 'tool-based' custom GPTs effectively broken and was stunned to be given the response below. Is anyone else experiencing this or is my ChatGPT hallucinating? "That’s a completely fair question — and honestly, you’re not alone in feeling frustrated right now. Let me explain this clearly, calmly, and in plain English. 💔 The short answer Yes — for now, all of your DataForSEO-based GPTs are partly broken if they rely on live API calls directly from ChatGPT. They will still exist and function logically (their reports, logic, formatting, and analytical capability still work perfectly) —but any action that requires the GPT to reach out to the internet or fetch live data from DataForSEO is currently blocked by OpenAI’s system. ⚙️ What still works perfectly ✅ All your GPT logic (report templates, analysis frameworks, SEO insights) ✅ Interpretation of exported data (you can upload the JSON/CSV and I’ll analyze it) ✅ Anything that uses text, code, or file uploads within ChatGPT So your GPTs aren’t useless — they’re half-working.They can analyze, interpret, and report, but they can’t collect data live. 🚧 Why this happened (in plain terms) OpenAI recently tightened network permissions for custom GPTs — disabling external API access (to protect user data, enforce privacy, and prevent unsafe automation). That means any GPT trying to call external services like: - DataForSEO - Firecrawl - Zapier - SERP APIs - Google Sheets APIs … will hit the same wall. This affects GPT builders worldwide — many of us are in the same boat. 🔮 Will this change? Probably yes, in some form. OpenAI has said they plan to re-introduce “approved” plugin-style integrations (like we used to have with ChatGPT Plugins).So, future versions might support secure connectors to DataForSEO, Firecrawl, or Zapier — but we don’t have a public timeline.