😺 Goldman Sachs' new AI job forecast & how to train your Claude (Code)...

Welcome, humans.

The results from last Friday’s GPT-5 poll are in, and surprisingly, y’all are pretty happy with GPT-5!

The top 3 answers:

1. GPT-5 = everything I wanted (297).
2. Team whatever works (207).
3. Missing GPT-4o’s personality (148).

Though to be fair, if you add up everyone who chose Grok, Gemini, or Claude over GPT-5, that would be the second highest answer at 275 votes.

Here’s what some of y’all said:

• First, the GPT-5 stans:
  • K.H: “The more I use GPT5 the more impressed I am with it. Been using it mainly for reasoning up to now. No coding yet. I'm a Plus subscriber and haven't gone back to 4o yet. Probably never will do.”
  • S.C.: “I asked it to expand a 6 paragraph plot to a 5000 word science fiction story and it totally blew me away with how good the writing was. Yes, I had to do some editing (obviously) but still... damn, I wish I could write like that!”
  • J.R: “GPT5 is clearly an upgrade. In every way. The over the top exuberance absolutely annoying. I like the calmer more pragmatic approach to answers, the more conversational vs celebrant tone, it’s much better at explaining itself and articulating ideas.”
• And from the GPT-4o die hards:
  • A.C: “4.0 was so much fun. I looked forward to every session. 5.0 is like a competent grad-school study buddy. Valuable, but no connection.”
  • I.D: “5 is truly better in coding and utility... 4o was a true partner in your projects.”
  • J.L (who chose team crisis averted): “I use ChatGPT primarily for writing and make heavy use of canvas docs. GPT-5 jacked up the way canvas docs function and it’s no longer efficient for me to use. Bye bye ChatGPT- Hello Claude Sonnet.” (Claude Squad unite!!)
• And everybody else:
  • L.N chose Team crisis averted: “Removing choice from the paying subscribers + no prior notice = poor user experience.”
  • A.C chose Whatever works: “Until GPT-5 came out, I was using Gemini 2.5 PRO almost exclusively. Now, I realize that GPT-5 is better than Gemini 2.5 PRO for MCP tool calling. So, I use them both throughout the life cycle of a research project.” (good tips!)
  • M.U chose Whatever works: “The SOTA seems to change almost weekly. I do not restrict myself to using one model, varying by the task that I need them for. Hallucinations, underlying bias and lack of ability to differentiate unreliable sources (e.g. marketing claims on a junk website) from reliable ones all make relying on these tools more risky than it would ideally be.”

For what it’s worth, OpenAI did post on X that GPT-5 should be warmer now with “no rise” in sycophancy. Perhaps we just got the best of both worlds?

Here’s what happened in AI today:

• Goldman Sachs forecasted AI will displace 6-7% of jobs, but only “temporarily.”
• Meta faced scrutiny after AI chatbot death incident and minor safety concerns.
• Google Search traffic to publishers dropped 10% in 8 weeks due to AI.
• Otter faced class action lawsuit over non-consensual call recording.

Advertise in The Neuron here

P.S: Do you know the best way to prompt your AI? What about when to use an agent vs a “project”? ICYMI, our new podcast episode covers these topics (and a lot more!). Listen and/or watch on… YouTube | Spotify | Apple Podcasts

While AI Reshapes Jobs, Smart Workers Are Reshaping AI…

Two stories caught our eye this weekend. First up, Goldman Sachs says AI will displace 6-7% of jobs (but create new ones), while AI educators are showing how to turn your computer into a personal assistant. Translation? The future belongs to people who use AI better, faster.

Goldman's take: Yes, AI will automate parts of jobs… especially for programmers, accountants, and customer support. But they predict only temporary unemployment before new roles emerge. Why? AI mostly compresses tasks within jobs, not entire roles. So stop thinking of your career in job descriptions, and start thinking of it in terms of job outcomes. Read our deep dive on this data (and what to do about it) here.

Speaking of outcomes… Claude Code just got stupid powerful. AI educator McKay Wrigley has a killer tutorial showing how to pair Claude Code with Obsidian so AI can read, write, and organize files directly on your computer. It can research using "sub-agents" and even work while you're away. Essentially, Claude Code is now the best personal agent available right now. Read our deep dive on how to leverage it here.

FROM OUR PARTNERS

Ditch print servers and slash helpdesk tickets by up to 95% with Vasion Print (formerly PrinterLogic), the secure, serverless solution for IT teams fed up with outdated legacy systems.

Manage users, devices, and workflows from one cloud-native platform designed with Zero Trust at its core.

Eliminate Print Servers for Good

Prompt Tip of the Day.

D-Squared distilled the top advice from OpenAI on how to prompt GPT-5 into these five key tricks you can apply to better “steer” the model. They are:

• Trigger Words (2:52): Use phrases like “thinking deeply,” “double check your work,” “be extremely thorough,” and “this is critical to get right” to increase the model's reasoning level
• Prompt Optimizer (3:33): Use OpenAI's prompt optimization tool (requires API credits, about $1-2) to automatically improve your prompts based on best practices.
• Words Matter/Be Specific (5:43): Avoid vague terms and contradictions; be extremely specific in your instructions (example: instead of “plan a nice party, make it fun but not too crazy,” say "plan a birthday party for my 8-year-old, 10 kids, $200 budget, 2 hours, unicorn themed")
• Structured Prompts (6:53): Use XML structure with tags like <context>, <task>, and <format> to help the AI better comprehend different sections of your system instructions.
• Self-Reflection (8:16): Have the AI create a rubric for itself based on your intent, then judge its own output against that rubric, iterating multiple times internally before giving you the final response.

Check out all of our prompt tips of the day from August here!

Treats to Try.

*Asterisk = from our partners. Advertise in The Neuron here.

1. *Spinach AI records, transcribes & summarizes your meetings, then automatically updates your CRM and project tools.
2. Higgsfield now lets you upscale videos and images.
3. Notion AI’s new AI tool will be able to use Notion across multi-step tasks like editing lots of pages or updating an entire database.
4. Continua joins your group chats to track important decisions, answer questions with live web search, and automatically turn your conversations into shared docs (raised $8M).
5. MolmoAct helps your robot understand and follow instructions by reasoning in 3D space, so you can tell it to “put the pillow on the couch” and it visualizes the path and movements needed to complete the task.
6. TrackingAI tracks the IQ tests of current AI models, while Artificial Analysis’ AI Trends page tracks models by architecture, open weights vs proprietary, efficiency, and top intelligence by country (between US and China).

Around the Horn.

• Google Search traffic is down 10% to “premium publishers” over the last 8 weeks, according to new data, while half of all journalists surveyed in select European countries now use AI in their writing process (paper).
• Meta has come under scrutiny for its AI chatbots after a Reuters investigation reported a cognitively impaired retiree died in a fall while rushing to meet “Big sis Billie” and revealed the company’s “Content Risk Standards” allowed bots to engage in “romantic or sensual” chats with minors—prompting Sen. Josh Hawley to open a probe the next day.
• OpenAI hosted a dinner to woo journalists ahead of a potential IPO (our invite must’ve been lost in the mail lol), hinting at new products it could launch, like an AI browser, AI social media app, and even a potential brain interface chip.
• Note-taking AI Otter was hit with a class action lawsuit that alleges its bot joins Zoom/Meet/Teams calls, records without all-party consent, and uses audio to train models; expect enterprise meeting tools to shift toward explicit all-participant consent prompts and clearer bot disclosures as a result.
• China hosted the first-ever World Humanoid Robot Games where robots from 16 countries competed in sports and practical tasks like customer service, as part of the country’s broader push to lead in AI development against the US.

FROM OUR PARTNERS

AI Agents That Cut Support Costs By Up To 80%

AI Agents Designed For Complex Customer Support

Maven AGI delivers enterprise-grade AI agents that autonomously resolve up to 93% of support inquiries, integrate with 100+ systems, and go live in days. Faster support. Lower costs. Happier customers. All without adding headcount.

👉 Book your demo to see how.

Monday Meme

A Cat’s Commentary.