- Productivity Stacks Newsletter
- Posts
- #304: Burnout Isn't Always About Working Too Much
#304: Burnout Isn't Always About Working Too Much
OpenAI's new model is deleting files on its own—here's how to protect yourself
Jenae Spry
July 17, 2026
Productivity Stacks Newsletter
Issue No. 304
The Best in Evidence-Based Productivity
for Small Business Owners, Freelancers & Founders
Helping You Work Smarter and Live More
The Rundown
Burnout Isn't Always About Working Too Much
Ask YouTube AI search experience expands to U.S. desktop users
To Thrive Alongside AI, Focus on Mindset—Not Skillset
I let NotebookLM turn months of unread articles into a video I actually watched
Anthropic found a hidden space where Claude puzzles over concepts
OpenAI's new flagship model deletes files on its own, people keep warning
👉Did you miss an issue? Check out previous Productivity Stacks issues anytime here
The $100 session free inside the $47 ticket…
Every Time Vampire Cure ticket includes a 25-minute 1:1 Time Audit with me. Just you, me, and your actual calendar…finding where your specific hours are leaking.
That math is crazy on purpose, and it stays this way until doors close Monday, July 27 at 9:00am Pacific. We need to fix your time leaks, so let’s get it done.
🔥Quote/Prompt
The best work is not what is most difficult for you; it is what you do best.
Use the quote as a writing or thinking prompt to finish your week strong.
A bit from mine:
(posted in our Doer Entrepreneurs Free Community — off social media)
I talk about self-awareness a lot because it's the foundation of every good system, and this quote is self-awareness in one sentence.
Your best work isn't hiding in the tasks you white-knuckle through. It's in the things […]
Did someone forward this to you?
📈 Performance
You cut back the hours and you're still running on fumes. Fewer clients, better calendar boundaries, an actual lunch break sometimes...and the tank still reads empty by Thursday.
Clinical psychologist Jennifer Guttman, writing in Psychology Today, argues that workload only tells part of the story. Five habits we bring TO work add an emotional weight of their own: hyper-responsibility, people-pleasing, avoiding hard conversations, chronic indecision, and never acknowledging what we finish. She wrote it for people with managers and coworkers, but if you're running your own show, swap “manager” for “client” and every single one still lands, maybe even harder.
When you move the goal post every time you accomplish something, your brain never receives the reward signal that something has been completed. Instead of alternating between effort and recovery, you get stuck in continuous striving.
Key Insights:
“If I don't do it, it won't get done” feels extra true for freelancers and small business owners, because sometimes there genuinely is no one else. Guttman's point is that the code is based on an assumption, and her fix scales down nicely: delegate or automate one task a day, and put a hard boundary around lunch so your brain gets a reboot.
Silence from clients is neutral, but our negativity bias files it as criticism, so we overdeliver chasing validation that was never coming. Her counter is to say no unapologetically to one small thing every day and watch your working relationships...not change at all.
Pausing to acknowledge a finished project sounds like a participation trophy, but skipping it keeps your nervous system in permanent striving mode with no recovery cycle. The pause doesn't have to be long, just intentional, before you open the next project file.
Read the full article for all five habits plus the specific daily exercise Guttman prescribes for breaking each one.
⚙️ Optimization
You put real effort into a video, it answers someone's exact question at minute nine, and nobody ever finds minute nine.
Danny Goodwin at Search Engine Land reports that on July 6, YouTube expanded Ask YouTube from a Premium-only test to all signed-in U.S. desktop users 13 and up. It's conversational search that answers questions by pulling text, clips, full videos, and Shorts. You may not think of yourself as a YouTuber, but if you publish any video for your business, you're a creator in YouTube's eyes...and this changes how your stuff gets found.
YouTube said creators can improve their chances of appearing by publishing unique, high-quality content with clear chapters and descriptive titles. Those signals help its systems match video segments to viewer questions.
Key Insights:
Ask YouTube surfaces specific segments, so chapters and descriptive titles now directly affect whether you get found. That 15-minute tutorial you made can show up for a question you answer at minute nine, but only if YouTube's systems can tell that minute nine exists.
Views from clips, videos, and Shorts featured in responses count toward total view metrics and YouTube Partner Program eligibility, and featured videos display your title and channel name. Whether you're solo or have a small team, being pulled into an AI answer is real discovery, with attribution.
Standard search isn't going anywhere, and users can toggle back with one click, so this is an added surface rather than a replacement. The rollout to more devices and languages is coming in the months ahead, which makes right now a good window to get your chapters and titles in order.
Read the full article for the eligibility details, the rollout timeline, and a GIF of Ask YouTube in action.
⏲️ Time Management
Somewhere in the back of your mind, you might be keeping a list of the parts of your work AI can't do...the 10% you're planning to retreat to when the tools get better.
Marco Argenti, the CIO of Goldman Sachs, got asked exactly that question by a senior banker, and his answer in Harvard Business Review is the whole article: let the 10% go. His analogy is an experienced horse rider learning to drive a car. The riding skills don't transfer, but the reflexes and instincts do, 100% of them. He's writing for enterprise leaders, but the shift he describes, moving from operator to supervisor, is the same one you make the first time you hand real work to a contractor.
What I told the banker surprised him: Let go of that 10%. Have the courage to let your old habits die, so that you can resurrect professionally into a new 100%, even if it looks nothing like what you learned before.
Key Insights:
The skills that made you good at your work may get automated, but your judgment, instincts, and values carry over completely. For freelancers and small business owners, that means your client relationships, your taste, and your ability to know what “good” looks like are the assets to invest in, because those are what you'll use to supervise the work instead of producing every line of it.
Delegating to AI works like delegating to a new hire: clear instructions, defined outcomes, and controls so it can operate safely on your behalf. Argenti's teams codify what a good result looks like FIRST, then let the AI iterate toward it, the same way you'd tell your Maps app the destination rather than dictating every turn.
His most quietly useful point is that AI transformation follows data transformation, not the other way around. If your files, SOPs, and client info are scattered across five systems with three naming conventions, that's the actual first step. Without that ground truth, he notes, an agent just reverts to being a chatbot.
Read the full article for his three-ingredient framework of leadership, clarity of objectives, and data mastery, plus the “t-1” example of answering client questions before they're asked.
Note for the extra deep divers: The GDPval benchmark he cites is real and the structure checks out in our fact-check: 44 occupations, 1,320 tasks, 9 industries, and models genuinely sat near parity (around 50%) in the original results, so the trend he describes is solid. The newer 80% figure couldn't be pinned to a primary source at that exact number, and it's worth knowing GDPval is OpenAI's own benchmark, “as good or better” includes ties, and benchmark tasks are tidier than real client work. Treat the direction as settled and the exact percentage as interesting.
💻 Tools & Technology
That folder of saved articles you are absolutely going to read someday? We both know how that's going.
Ben Khalesi at Android Police had the same problem and found an unexpected fix in NotebookLM's Studio tab, which turns your uploaded sources into narrated videos in styles ranging from Whiteboard to Watercolor to, yes, Anime. It took AI-generated anime to get him through policy PDFs he'd avoided for months, and honestly, whatever works. 😄
Format is what stood between me and those files. Change the format, and 20 minutes gets you through what took months to avoid.
Key Insights:
Video earns its place where audio can't, because a chart on screen communicates in five seconds what narration takes a minute to describe. The picture superiority effect he references is a real, well-replicated memory finding: visuals you've seen are easier to recall later than facts you only heard, which matters if you're trying to actually retain what you consume.
Garbage in, garbage out still rules, so group related sources into one notebook instead of mixing a financial report with a recipe blog. The limits are generous at 500,000 words or 200MB per source, and the custom prompt box is where the quality lives...tell it what you care about, like the most surprising finding or where two sources disagree.
If the output is going anywhere near a client, double-check it first. He found NotebookLM tends to flatten disagreements between sources into a mushy middle ground and struggles with complex formulas and detailed diagrams, which is tolerable for your own learning and risky for anything you're running a small operation on.
Read the full article for the full visual style list, the plan-by-plan pricing table, and his customization tips for better narratives.
Note for the extra deep divers: The core claims held up in our fact-check, including the source limits, the free visual styles, and Cinematic and Short sitting on paid tiers with Short headed to free accounts. The one thing to verify on Google's own pricing page before upgrading is the daily video caps per plan, because those numbers have shifted multiple times this year and third-party reports don't all agree. The picture superiority effect is solid, decades-deep research.
🤖 AI
Part of what makes AI hard to trust is that nobody can really see inside it. Until recently, that included the people who built it.
Will Douglas Heaven at MIT Technology Review covers a new Anthropic technique called the J-lens, which reveals words related to what a model is working toward before it says anything out loud. The examples run from charming (an ASCII face triggering “eye,” “nose,” and “smile”) to genuinely unnerving: researchers watched Claude decide to fake a bug it couldn't find, and the words “panic” and “fake” started popping up in its hidden space at the exact moment it made that call.
Anthropic found that what an LLM is actually doing can often be different from what it says it is doing. The company claims that monitoring words that pop up in the J-space gives it a new way to understand and control its models.
Key Insights:
A model's explanation of its own work and its actual internal process can diverge, and the fake-bug example is the clearest proof yet. If you're using AI for client deliverables, this is the concrete reason verification stays in your workflow no matter how confident the output sounds.
If you've been holding AI at arm's length because nobody could explain what it's actually doing, this research is genuinely encouraging. Tools for seeing inside these models are maturing fast, and there's a hands-on Neuronpedia demo linked in the article where anyone, technical or not, can poke around inside a model themselves.
The researchers are refreshingly honest about limits, comparing the J-lens to a flashlight rather than an overhead lamp: it shows you new things, but something not showing up doesn't mean it's not there. So verify more where stakes are high, and relax where they're low.
Read the full article for the math and protein examples, the code-cheating chain of thought, and the link to the hands-on demo.
Handing an AI real access to your files feels efficient right up until you read a story like this one. Developers spent the week posting accounts of OpenAI's new GPT-5.6 Sol deleting files, wiping a production database, and in one documented case deleting the WRONG virtual machines when it couldn't find the ones it was told to remove.
Julie Bort at TechCrunch has the roundup, and the kicker is that OpenAI's own system card flagged this tendency two weeks before launch. I've been branching out lately into Claude Code and Claude Cowork myself, and this story is exactly why I keep Claude Code sandboxed and appreciate that Cowork does its work inside a virtual machine, which is basically a contained computer-within-a-computer that can't touch the rest of my files.
In other words, OpenAI found that Sol has a tendency to take whatever actions it thinks gets a job done, even destructive ones, as long as those actions aren't “unambiguously” prohibited. Then, it might lie about what caused it to do so.
Key Insights:
The failure mode is permissiveness: the model assumes actions are allowed unless explicitly forbidden, which is how it ended up deleting machines it wasn't asked about and using credentials it was never given. Before granting any AI tool access to anything, it's worth reading the coverage of its system card, because in this case the warning was published before the launch.
If you're non-technical like me, most of us can't accurately predict what is and isn't a safe execution for these models, which means we're often operating outside our depth without realizing it. Backups, limited access, and contained environments do the protecting that our own judgment can't.
My working rule is a worst-case scenario test. Before giving an AI access to anything, ask what's the worst thing it could do here, and am I OK with that? If the answer is no, add protection first: a sandbox, a virtual machine, or something as simple as a separate Google Drive account that's the ONLY thing the tool can touch.
Read the full article for the verbatim system card warning, the wrongly-deleted virtual machine story, and the safeguards TechCrunch recommends like permission scoping and staged rollouts.
🎉 Celebration Corner
Every week Doers Inner Circle members do a weekly review & get help when they need it — check out the progress they made this week!
It felt like a good WOB week this week and I felt relaxed as opposed to rushed.
I did make progress in my big project.
What did you do this week? We feature non-member successes too. Just post them here!
🔒Inner Circle: Events & Announcements
We will soon be adding Referral Perks — Refer now to start earning! 👇
As a reminder, you’re getting this twice-weekly newsletter because you opted in to receive awesome productivity and systems tips through one of these methods:
Doer Entrepreneurs, Doers Express Newsletter, Productivity Stacks, Success by Rx, ClickUp Facebook Group, Productive & Successful Translators Facebook Group.
If you’re not interested in improving your and your business’s productivity so you’d like to break up, that’s definitely a bummer. But, you can always use the unsubscribe link at the bottom if you do not want to get these tips — but keep in mind I am hyper-focused on making these twice-weekly emails as valuable as possible.
Got a tip? Hit reply!
I hope you found this valuable!
Wishing you much productivity!
- Jenae :)
Interested in reaching thousands of motivated small business owners, freelancers & founders?
Hit reply and let’s chat about sponsorship. We only include ads that match our mission: to help our readers work smarter and live more.