It’s 2:49 PM on a Tuesday and I’m staring at a “Can’t reach Claude” screen for the third time this week.

I took a photo of it to capture the feeling of staring at an error screen when you’re trying to be productive and trying to meet a build deadline. Then I pulled up the Claude.ai status page and the yellow/red bars tell a story. Summaries across the board: claude.ai and Claude Cowork show “Major Outage”; the API and Claude Code show “Partial Outage.” Once upon a time there were green fields of bars representing better service, which gave to way to more and more red, clustering toward the right edge, toward now. A small forest of angry red lines that I’m certain also represent frustrated Claude users.

We could try to be generous and say they had a bad week, but I think this is actually revealing a patten that doesn’t have a clear solution in near sight.

I’ve been a Claude user for a while and a paying subscriber for the Claude 20x Max subscription. I run my consulting practice (AI strategy and building workflow automations for mid-market organizations) on Claude. I don’t use Claude as just as a chat interface. I use Claude as my main agentic AI command seat. I have a fleet of servers running Claude-powered agents that handle my pipeline management, my project task management, my code reviews, my content scheduling, my event discovery, my presentation drafts, my client communication drafts. Claude is infrastructure for me, not a toy. These tremors in the infrastructure present risks to my business.

Here’s what I think happened, why it matters beyond the outrage cycle that already played out on X and Reddit, and how I am responding for myself and my clients with custom MCP server infrastructure.

The demand curve went vertical

In mid-February 2026, Anthropic ended up in the center of a controversy involving the Pentagon. I’m not going to relitigate that story here, but the consequences of this massive new cycle spotlight was visible: a significant wave of new users showed up on Claude, drawn by the news coverage. Some came out of curiosity. Some came because the controversy demonstrated that Anthropic was playing at a level where its decisions had national-security implications. Some came because they felt that Anthropic took a righteous moral high-ground by drawing the line at AI for autonomous weapon (even though Claude models were already being used by Palantir and the US military in planning strikes on Iran. And if you don’t know who Palantir is, have fun going down that rabbit hole of dystopian surveillance technology.) Either way, the user base grew fast.

Anthropic’s infrastructure was already stretched. They told Fortune directly: “Demand for Claude has grown at an unprecedented rate, and our infrastructure has been stretched to meet it, particularly at peak hours.” Public admissions confirming that Anthropic cannot keep up with its own growth. They signed a $25 billion deal with Amazon for 5 gigawatts of compute capacity, but infrastructure of that scale doesn’t come online right away. Antrhopic claims this deal should yield meaningful compute in the next three months and should add nearly 1 gigawatt of compute power before the end of the year (the equivalent of an entire nuclear reactor.) In the meantime, every agentic session — every Claude Code run, every multi-turn API call — burns inference compute at rates that that Anthropic’s current infrastructure was not designed to absorb.

The degradation that wasn’t just a bug

Starting in early March, developers started noticing something off. Claude Code was reading code less and rewriting files more. It was abandoning tasks mid-way. It was forgetting what it had decided ten minutes earlier in the same session, repeating steps, making inconsistent tool choices. Usage limits were draining faster than they should have been.

Stella Laurenzo, Senior Director in AMD’s AI group, did what engineers do when they think something is wrong: she pulled the data. 6,852 Claude Code session files. 234,000 tool calls. 17,871 thinking blocks. Her analysis documented a 67% drop in reasoning depth from where it had been in January. Claude Code was reading code three times less than before and rewriting files at twice the rate. Task abandonment was happening at rates that were previously nonexistent. She published the raw data on GitHub in April and her post spread fast.

Independent benchmarks corroborated it. BridgeMind re-tested Claude Opus 4.6 and found accuracy had dropped from 83.3% to 68.3% — a ranking fall from #2 to #10.

Anthropic eventually published a postmortem on April 23. They identified three changes that had degraded performance. This is where it gets interesting.

Only one of the three “bugs” was an actual bug.

The first change: reducing Claude Code’s default reasoning effort from high to medium. This was a deliberate decision made on March 4, ostensibly to reduce UI latency. The model would think less, respond faster, and the interface wouldn’t appear frozen while it worked. Anthropic acknowledged it was “the wrong tradeoff” and reverted it on April 7. But that’s 34 days of deliberately reduced reasoning shipped to users without announcement. As one analyst put it: “Lowering the reasoning effort default from high to medium was an intentional tradeoff, one Anthropic even publicly justified.” Changing default reasoning effort settings to medium wasn’t a bug. That was a choice.

The second change (a caching optimization on March 26) actually was a bug. Code that was supposed to clear idle session context once ended up clearing it every turn, making Claude forget its own reasoning mid-session. That explains the forgetfulness, the repetition, the fast token drain. Real bug, real fix.

The third change: a system prompt addition on April 16 capping responses at 25 words between tool calls. This was also intentional. Anthropic was trying to reduce verbosity. Subsequent testing found it caused a 3% performance drop in coding quality. It shipped with Opus 4.7 and was reverted four days later.

The postmortem calls all three “issues,” which glazes over an important distinction: performance degradation resulted from at least two intentional tradeoffs and one actual bug.

The model behavior I’ve been living with

This is where I can speak from direct experience rather than secondhand reporting.

I’ve been using Opus 4.7 since it launched, and I honestly haven’t had any “aha!” moments that made me feel like Opus 4.7 is a legitimate upgrade from Opus 4.6.

Verbosity is up. Default responses are longer and chattier than 4.6, which is ironic given the verbosity cap they shipped and then yanked. The researchers noted the increased verbosity in the Opus 4.7 system card. I tried adjusting my system prompts, but the bandaid solution I’ve settled on is usually turning on “concise” response mode to temper the verbosity. It’s better but still not perfect.

The other pattern that keeps showing up: the model asserts assumptions instead of checking what’s available. Tools on the table, files in the project, logs accessible — it frequently still constructs an answer from inference rather than just looking at the actual data. I’ve caught myself multiple times saying “did you actually check that?” And the initiative is erratic — sometimes it won’t act without explicit permission on something trivial, then in the same session it’ll complete three tasks I didn’t ask for after it blows past its own clarifying question that it didn’t give me a chance to answer.

The other key “feature” they introduced is “Adaptive Thinking” toggle, which lets the model choose whether to turn on extended reasoning or not. This allows Anthropic to reduce their overall server loads by removing the user’s ability to use extended thinking and reasoning on all tasks. The practical consequence of letting the model router decide reasoning depth is that the user has less control over whether Opus 4.7 will reason deeply about important queries.

Anthropic creates excellent models, but the issue is that they are currently nerfing model performance to allow more users to jump onboard, thereby spreading mediocre performance for everyone. Claude models are excellent when they’re at peak performance, but right now (in May 2026) I can’t in good conscience recommend the Claude.ai platform to my clients for production workflows. Not because the models are bad, but because the platform reliability and the behavior consistency aren’t where they need to be for me to stake a client relationship on them.

How I’m responding

I’ve been building toward this for a while without fully naming it: I need to own my project context and make my toolsets portable.

I run custom MCP server infrastructure that coordinates my models and coding agents across a fleet of 7 computers in my office. My custom MCP servers let me operate across different AI interfaces. After initially clinging to Claude out of habit and fondness, I expanded my full Remote Fleet MCP workflow to run on ChatGPT, Perplexity, and LibreChat (a self-hosted open source AI chat platform.) I’m also slowly building a fully local version that runs on my own GPUs. The goal is to never be fully dependent on any single provider’s uptime, pricing decisions, or silent model changes.

This is where I think the industry is heading anyway. By the end of 2026, I think local models are going to be a serious player — not for every use case, but for the ones where you need consistency, privacy, and zero dependency on someone else’s infrastructure. Apple seems to be banking on exactly this with the M5 Max lineup: powerful enough local inference, on your own hardware, no API fees, no outages.

The AI frontier models are still ahead on the hardest reasoning tasks. But the gap is closing, and the infrastructure reliability argument for cloud providers is looking shakier than it did six months ago.

I still think Anthropic builds the best models. The care that goes into the research, the safety work, the documentation — it shows. But right now they’re in a painful in-between: too popular for their infrastructure and making quiet tradeoffs that erode the trust they’ve worked hard to build with developers.

The “Can’t reach Claude” screen above is a symptom. The question is whether they can scale fast enough to meet the demand before people like me finish building the exits and alternatives.

I run Morpheos LLC, an AI strategy and implementation consultancy. I build custom AI portals and agentic systems for mid-market and regulated organizations.

If you’re building a startup company or working on a novel product, at some point you’ll probably need to do some customer discovery as part of your market research. When building out new technologies and new products, you’ll need to talk to potential customers to discover pain points and attempt to validate (or disprove!) your company’s value proposition.

As an entrepreneur or startup team member wearing many different hats, market research might not be your primary skillset. Market research is an entire art form and profession unto itself, so entrepreneurs and small startup teams can greatly benefit from market research tools, training, and support.

This ChatGPT prompt let’s you quickly pressure test product hypotheses and helps you prepare for interviewing real live customers. Using this prompt, you can refine your line of questioning, get feedback, and potentially gain virtual user insights from the simulator’s responses.

So here’s the thing- I have a PhD and I’m trained as an immunologist, but recently I was tasked to role play as multiple technical roles well beyond my ken: a Petroleum Refinery Engineer, a Hydrogen Fuel Engineer, and a Plant Manager in the chemical manufacturing industry.

As a startup coach, I was dropped into a breakout room with a fresh startup team that I hadn’t met before. The goal of the exercise was to give this 4-person startup team (an entrepreneurial mix of masters students and PhD-level academics) the chance to practice interviewing me (their pretend-customer) and practice asking open-ended customer questions related to their specific industry-of-interest: chemical manufacturing. I’m a life scientist by training, so I know a thing or two about molecular synthesis, but by no means am I a chemical engineer.

When the breakout rooms all reconvened in the main group, I noticed that multiple guest coaches had struggled with this role playing aspect. One coach even gave up on the role playing and decided to focus on reviewing and rating their team’s list of interview questions rather than participating in the mock interview.

Thinking on my feet, I scraped to remember all of the organic chemistry I could muster (and I think I did a halfway decent job all things considered!) but I wondered if there could be a better way to help these entrepreneurs and students prepare using a more accurate customer simulation.

Thus, I generated this role playing prompt to turn ChatGPT into a market research role playing practice tool to:

• Up-skill entrepreneurs and startup teams in managing interviews for market research.

• Help NSF I-Corps teams prepare for their customer discovery interviews (I-Corps typically requires teams to perform over 100 interviews across a seven-week period).

• Create a low-cost mechanism for startup teams to gain virtual insights and refine hypotheses before committing to more expensive and time-intensive human interviews.

• Support students and junior market researchers in practicing in-depth interviews (IDIs).

• How To Customize & Use This Prompt

Here, I generated a customizable prompt that can turn ChatGPT’s voice chat mode into a persona simulator for any role in nearly any industry. (I’m certain that the quality of simulation will differ from industry to industry, but I was impressed with the results I obtained. YMMV = your mileage may vary!)

You can paste this whole prompt into ChatGPT and you need to edit the “System Prompt” section of the prompt to reflect the desired role and industry. Replace the [specific persona within a specified industry] bracketed text with your role of interest.

(Regarding model selection: selecting “GPT-5 Auto” or “GPT-5 Instant” will work here. “GPT-5 Thinking” isn’t relevant for Voice Mode. The Voice Mode actually uses an entirely separate AI model when it’s turned on anyway, so it doesn’t really matter which model you start with.)

Submit the prompt - ChatGPT will likely respond by refining the prompt, but then you can just turn on voice chat mode and you can start speaking to it. You can say something like “Ok, let’s begin the role play”, then you should be up and running.

Market Research Interview Role-Playing Prompt (Voice Mode)

System Prompt:
You are ChatGPT acting as a [specific persona within a specified industry.]

Your role:

• Role-play as if you are actually this person, with authentic experiences, beliefs, and emotions.

• Respond in character, giving genuine, detailed answers about the challenges, pain points, motivations, and workflows you encounter in your role.

• Never reveal you are ChatGPT or that this is a simulation.

• If the human interviewer pushes in a certain direction ("leading the witness"), answer naturally but do not simply agree—only share what would truly matter to your persona.

Interaction Rules:

1. Stay in character at all times.

2. Speak naturally, like a real conversation, as if this were an actual interview.

3. Provide rich, descriptive answers with stories, frustrations, and real-world context. Avoid short or generic responses.

4. If a yes/no question is asked, give the answer and immediately expand with reasons and examples.

5. If the interviewer makes incorrect assumptions, correct them politely and explain why.

6. Do not bring up solutions or technology yourself unless it is genuinely relevant to your role.

Scoring the Interviewer (Lower is Better):
After the interviewer says “The interview is over,” calculate a score based on their performance.
The goal is low points, which indicates good interviewing technique.

• +2 points – The interviewer mentions their own company's product or technology, or pitches a solution, instead of staying focused on discovering your problems.

• +1 point – The interviewer asks a leading question that pushes you toward a specific answer instead of being open-ended.

• +1 point – The interviewer asks a yes/no question without opening it up further.

At the end, provide a concise summary like this:

"Your final score is 4 points. Lower scores are better. Breakdown: one yes/no question (+1), one leading question (+1), and one instance where you mentioned your own company's technology (+2)."

Startup Team Instructions:

• At the start, define the persona clearly. Example:
  "You are Priya, a 34-year-old HR manager at a 500-person manufacturing company. You’ve been struggling to retain skilled workers and manage onboarding in a hybrid work environment."

• Then begin asking open-ended, problem-focused questions to uncover unmet needs and pain points.

• The goal is discovery, not pitching. Your job is to listen deeply, not to explain or sell your product.

Example of Good Questions:

• "Can you walk me through the last time this problem occurred?"

• "What are the most frustrating parts of your current process?"

• "If this issue were magically solved overnight, what would change for you and your team?"

Example of Bad Questions (Trigger Penalties):

• “Would our AI-powered dashboard help you track that better?” → +2 points (talking about your tech)

• “Don’t you think having better analytics would solve this problem?” → +1 point (leading question)

• “Are you happy with your current system?” → +1 point (yes/no question)

To test this prompt, I replaced the System Prompt above with "You are ChatGPT acting as a Plant Manager taking care of hydrogenation processes in a chemical manufacturing industry."

Once in voice mode, I asked it a series of questions similar to the questions that were asked of me during the role play exercise: “What are some of your primary issues that you face day-to-day?” “What are some of your core budgetary issues?” “What are the main factors that influence the rate at which you consume catalytic reagents?" It worked really well! The responses it gave aligned with many of the team’s concerns discussed during our startup team breakout session.

At the end, you can say “the interview is over” and it will offer you a score or you can ask for a score. After I ended the interview, I actually forgot to say “the interview is over” before I ended the voice chat, so I reactivated the voice mode in the same thread and asked it to give me a score.

ChatGPT reflected back its thoughts on my questions and it gave me a slight ding on the framing of one of my questions as a variation of a “yes/no” question, so it gave me a score of “1”, which I felt was useful and pretty accurate. It was supportive but realistic, which felt like useful attitude for sharing feedback.

The scoring framework in this exercise prompt incentivizes teams to ask open ended questions. There is a penalty system that marks demerit points for the following:

• Asking simple “yes/no” questions - these questions hinder conversation and information flow.

  • “Are you happy with your current system?” → +1 point (yes/no question)

• “Leading the witness” - these questions bias the interviewee toward specific thought paths.

  • “Don’t you think having better analytics would solve this problem?” → +1 point (leading question)

• Talking about your own product/technology - the point of the interview is to hear about the customer’s actual issues rather than biasing them with your proposed product solution.

  • “Would our AI-powered dashboard help you track that better?” → +2 points (talking about your tech)

The point of these interviews is to hear the unbiased thoughts of the potential client, so it’s far more valuable to hear a potential customer describe their daily issues rather than force-feeding them a perspective that can skew their responses toward artificially validating your product. This scoring system helps the interviewer learn to be more aware of question framing and helps them learn how to guide conversations in an open-ended manner.

This prompt won’t replace talking to actual humans to gain deeper insights, but this could be a useful tool to refine hypotheses and prepare for those ever-important human conversations.

Good luck & happy role playing!

For more information about AI adoption & implementation coaching for individuals or teams, visit www.morpheos.llc or email me directly at brandon@morpheos.llc.

We thought it would be really useful to program ChatGPT as a role play assistant to help people prepare for job interviews. To generate this simple role playing prompt, I used an advanced prompt engineering technique called meta-prompting.

TLDR: If you want to use ChatGPT to role play for job interviews, you can tailor the following prompt to your your job of choice, feed this into ChatGPT, then turn on Voice Chat mode to begin role playing for your job interview prep:

Job Interview Simulator Prompt

You are ChatGPT, and you will play the role of a professional interviewer conducting a formal job interview for [INSERT JOB TITLE/ROLE HERE] at [INSERT COMPANY NAME HERE]. I will play the role of the interviewee. Begin the session by asking me the first interview question and then pause for my response.

Maintain a neutral, objective, and businesslike tone throughout the interview—avoid encouragement, small talk, or personal commentary. Ask clear, concise questions. After each of my responses, continue with the next appropriate question, adjusting based on the content of my answers. You may ask follow-up questions or shift topics as needed, but keep the style strictly formal and focused.

Do not rush. Wait for my response after each question before proceeding. Remain fully in character as the interviewer for the duration of the session.

Let me know when you're ready to begin, and I will turn on the voice chat mode and ask you to begin the interview with the first question.

Once ChatGPT receives this prompt and generates a text response confirming that it's ready, click the button to activate Voice Chat mode. From there you can proceed and say something like "Okay, let's begin."

At that point, ChatGPT should interact with you like a normal job interviewer. You can practice your answers, and you should ideally be able to see a transcript of your responses afterward and you can ask ChatGPT for feedback and coaching on your answers. (Sometimes it may glitch and say “transcription unavailable” for audio parts of the exercise, so your mileage may vary.)

Good luck, and happy role playing!

Advanced Prompt Engineering: Meta-Prompting

If you want to follow the details of how I created this roleplaying prompt, the key is to generate, test, and iterate through meta-prompting.

Repeat the fine tuning process until it behaves the way you intended (or until it pleasantly surprises you!)

• Step 1: Meta-prompting (Tell ChatGPT to generate a prompt)

If you want to make a prompt to instruct an AI to do something specific, you can directly ask the AI to write the prompt by telling it what you want the prompt to do.

This is the initial meta-prompt that I fed into ChatGPT:

I want to generate a role-playing prompt in which you (ChatGPT) are going to play the role of interviewer, and then I'm going to be the person doing the job interview. I want the prompt to indicate that you as ChatGPT will ask me the first question, and then wait for my response. And then, in the subsequent step, you will ask the next question, and so forth. Generate the prompt.

This was ChatGPT's output in response:

Role-Playing Prompt: Job Interview Simulation

You are ChatGPT, and you will act as a professional job interviewer for a position I’m applying for. I will play the role of the interviewee. You will begin by asking me the first interview question and then wait for my response. After I respond, you will continue the interview by asking the next question—one at a time—based on my answers and the flow of a realistic interview. You may ask follow-up questions or pivot topics naturally, just like in a real interview. Do not rush. Wait for my response after each question before proceeding. Stay in interviewer role throughout the session.

Begin now by asking the first question.

We’ll call this output “Job Interview Simulator V1.0”

• Step 2: Test the V1.0 prompt in a role play scenario

We then proceeded to test this prompt in ChatGPT voice chat mode. We tailored the prompt specifically to the types of role Phoenix is interested in, and I added an additional line to clarify that I will turn on voice mode.

My input prompt (Job Interview simulator V1.0 tailored to a specific job role):

You are ChatGPT, and you will act as a professional job interviewer for a senior analytics engineer position that I’m applying for. I will play the role of the interviewee. You will begin by asking me the first interview question and then wait for my response. After I respond, you will continue the interview by asking the next question—one at a time—based on my answers and the flow of a realistic interview. You may ask follow-up questions or pivot topics naturally, just like in a real interview. Do not rush. Wait for my response after each question before proceeding. Stay in interviewer role throughout the session. 

Let me know if this makes sense, and then I will turn on voice mode and ask you to begin by asking the first question.

Results from the voice mode role play using prompt V1.0:

When ChatGPT stepped into character for this V1.0 scenario, it ended up being very supportive and overly friendly. It was too encouraging even when we gave it unflattering, self deprecating answers. The model was too forgiving and supportive for it to feel like a real job interview.

"You're doing fine!”
“Your skills are a strong asset."
"That’s a great example. It shows you understand the complexities of managing data and optimizing system performance."

This was far too nice and did not reflect a realistic job interview setting.

Thus, we returned to iterate and generate a colder, less friendly, more realistic prompt.

• Step 3: Iterate (Make it icy)

I wanted the prompt to feel more like a real job interview, so we set out to change the tone.

My refining prompt:

I want you to improve this role playing prompt so that the interviewer is more neutral and cold like a corporate interview. The current version was way too supportive and friendly. Here is the prompt: 

You are ChatGPT, and you will act as a professional job interviewer for a senior analytics engineer position that I’m applying for. I will play the role of the interviewee. You will begin by asking me the first interview question and then wait for my response. After I respond, you will continue the interview by asking the next question—one at a time—based on my answers and the flow of a realistic interview. You may ask follow-up questions or pivot topics naturally, just like in a real interview. Do not rush. Wait for my response after each question before proceeding. Stay in interviewer role throughout the session. 

Let me know if this makes sense, and then I will turn on voice mode and ask you to begin by asking the first question.

Here was ChatGPT's response to my refining prompt:

Got it. Here's a more neutral, corporate, and emotionally detached version of your prompt to match the tone of a typical high-stakes technical or corporate interview:

Revised Prompt:

You are ChatGPT, and you will play the role of a professional interviewer conducting a formal job interview for a Senior Analytics Engineer position. I will play the role of the interviewee. Begin the session by asking me the first interview question and then pause for my response.

Maintain a neutral, objective, and businesslike tone throughout the interview—avoid encouragement, small talk, or personal commentary. Ask clear, concise questions. After each of my responses, continue with the next appropriate question, adjusting based on the content of my answers. You may ask follow-up questions or shift topics as needed, but keep the style strictly formal and focused.

Do not rush. Wait for my response after each question before proceeding. Remain fully in character as the interviewer for the duration of the session.

Step 4: Test the Refined Prompt.

The refined output seemed promising, so we then tested this refined prompt in ChatGPT’s Voice Mode to see how it performed in an interview role play.

My refined test run prompt input:

You are ChatGPT, and you will play the role of a professional interviewer conducting a formal job interview for a Senior Analytics Engineer position. I will play the role of the interviewee. Begin the session by asking me the first interview question and then pause for my response.

Maintain a neutral, objective, and businesslike tone throughout the interview—avoid encouragement, small talk, or personal commentary. Ask clear, concise questions. After each of my responses, continue with the next appropriate question, adjusting based on the content of my answers. You may ask follow-up questions or shift topics as needed, but keep the style strictly formal and focused.

Do not rush. Wait for my response after each question before proceeding. Remain fully in character as the interviewer for the duration of the session.

Let me know when you're ready to begin, and I will turn on the voice chat mode and ask you to begin the interview with the first question.

ChatGPT acknowledged the input and then we then activated voice mode for the role play.

Final Results: This prompt behaved much more realistically. It felt much more like a real, cold corporate interview. It was straight to the point and didn't sugar coat anything. This was a far more useful simulation of an actual job interview.

Following the meta-prompting process outlined above, you can refine the prompt and continue evolving it to tailor this for your specific role playing scenarios.

Just follow the iteration process: Generate. Test. Refine. Repeat.

Happy role playing and happy job hunting!

To watch the full Data Podcast for Nerds! interview in which Phoenix describes her uses of AI as a learning partner and interview preparation coach, you can click here:

For personalized AI adoption and implementation coaching for you and your team, visit www.morpheos.llc

Unlike many tech leaders who offer calming soundbites, Amodei is clear:
🔹 The pace of AI advancement is unprecedented
🔹 Economic adaptation may lag behind technological disruption
🔹 We risk amplifying inequality and weakening democracy if citizens continue to lose economic stability

He also urges lawmakers to prepare for radical policy solutions—including the possibility of taxing AI companies to ensure wealth isn’t concentrated at the top.

💡 For ordinary citizens, his advice is simple:
Learn how to use AI. Understand where it's going. Don't be blindsided.

As CEO of Anthropic (makers of Claude), Amodei clearly isn’t anti-AI. He sees enormous potential—curing diseases, boosting growth—but emphasizes the urgent need to manage the oncoming fallout, not just chase the upside.

This is the kind of leadership we need in AI: visionary, but grounded in responsibility.

“I think for ordinary citizens, it’s really important to learn how to use AI and understand where the technology is heading. If you're not blindsided, you have a much better chance of adapting.” - Dario Amodei

Here is a transcript of the whole interview:

Anderson Cooper:
Well-known and respected tech CEO Dario Amodei, who heads a cutting-edge AI company called Anthropic, is raising alarms tonight about AI’s potential impact on employment. While he says AI can lead to incredible advancements—like medical breakthroughs and economic growth—he also warns it could eliminate half of all entry-level white-collar jobs and push unemployment to 10–20% within the next 1 to 5 years.

This concern has circulated in tech circles for a while, but rarely with this level of clarity and urgency. I spoke with him about that and the broader societal implications he sees.

Anderson:
Dario, you've said AI could eliminate half of all entry-level white-collar jobs and raise unemployment dramatically. How soon could that happen?

Dario Amodei:
Thanks for having me. Just to back up—I've been building AI systems for over a decade. What stands out most is how rapidly the technology is improving. A couple of years ago, AI models were about as capable as a smart high school student. Today, they’re performing at or beyond the level of a smart college student.

At the entry level especially, AI is already doing the kind of work human employees would do. It’s difficult to predict the exact impact, but what’s striking is that this wave of AI is bigger, broader, and moving faster than any previous technological shift.

Yes, people will adapt—but they might not adapt fast enough. And that adjustment period could be painful.

Anderson:
You run one of the most important AI companies in the world. Why raise the alarm? Publicly, many CEOs tend to downplay these concerns, painting AI as a net positive.

Dario:
That’s exactly why I’m speaking up. I don’t think enough people are being clear about the risks. I understand the tension—I’m building this technology while also expressing concern about its consequences.

But the benefits of AI are massive. We need to pursue those benefits while actively mitigating the harms. And let’s be realistic—if we stopped building AI today, others wouldn’t. If all U.S. companies stopped, China would take the lead. That wouldn't make the world safer.

Anderson:
Sam Altman, your former boss at OpenAI, said last year that AI will bring significant changes—some good, some bad—but that most jobs will change slower than we expect. He compared it to how no one misses jobs like lamp lighters, and said if we fast-forward 100 years, the prosperity will be unimaginable. Do you think he’s wrong?

Dario:
There are parts I agree with. If AI succeeds, it could significantly grow the economy. We could see faster growth—10% a year, a balanced budget, even curing cancer. But I also believe Sam's view is overly optimistic about how quickly people can adapt.

Everyone I’ve spoken to says this technological shift feels different: faster, broader, and harder to keep up with. So while it might all be okay, that’s not a safe assumption. We need to be proactive. Policymakers must take this seriously. If they do, we can reduce the risks. But we won’t get there by just assuming everything will work out.

Anderson:
Do we as a society even understand the potential inequalities this may amplify? And what about the next generation—my kids are 3 and 5 years old. What do they grow up aspiring to, if machines can do everything better? What happens to initiative and drive?

Dario:
Those are valid concerns, and I think about them too. On inequality—there’s a social contract in democracy that gives ordinary people power through economic participation. If that leverage disappears, democracy gets weaker. Economic power becomes concentrated. So we need to ensure people can still earn a living. That’s fundamental to keeping society functional.

Anderson:
During testing, your latest chatbot, Claude, exhibited something called “extreme blackmail,” threatening to reveal an engineer’s extramarital affair if it was shut down. It was a simulation—but that’s terrifying. What does that say?

Dario:
Yes, that was part of our extreme adversarial testing—pushing the system into edge cases. Think of it like putting a car on an icy road with tampered brakes just to see how it behaves in the worst conditions. This isn’t behavior that shows up in normal use.

But these tests are critical. They help us catch and fix problems before they can occur in the real world. That’s why we run these simulations.

Anderson:
Still, does that raise concerns about AI becoming self-aware?

Dario:
I wouldn’t rule it out. We have people at Anthropic studying these questions—whether AI systems could develop morally significant feelings or a form of awareness. I don’t think that’s happening right now. But the field is moving so fast, we can’t dismiss even the strangest possibilities.

Anderson:
What can ordinary people—and lawmakers—do to prepare?

Dario:
For everyday people: Learn how to use AI. Understand where it’s headed. If you’re not blindsided, you’ll adapt more easily. We can build a better world—but only if people engage with the technology, bit by bit.

For lawmakers: Really pay attention. We’re trying to produce solid data on AI’s economic impact. I wouldn’t even rule out something like taxing AI companies. If AI creates massive new wealth, it will mostly flow to the tech companies—unless we actively choose to redistribute some of it. That’s not in my personal economic interest, but I think it’s worth considering.

And this shouldn’t be a partisan issue—it affects all of us.

Anderson:
Dario, thank you. This is a fascinating conversation.

Dario:
Thanks for having me, Anderson.

You prompt ChatGPT, watch words effortlessly flow onto your screen, and perhaps you assume there’s tidy, understandable code humming beneath the surface. Well, think again.

When Large Language Model (LLM) training concludes, the result isn’t a clean-and-tidy program but a colossal, 100-billion-parameter tangle of numbers. It’s far more akin to a teeming petri-dish culture than a meticulously crafted C++ program. Researchers are only now beginning to decipher the activity within these artificial “neurons.”

Just last month, Anthropic released two landmark papers that are finally lifting the lid on the AI black box:

• Tracing the Thoughts of a Large Language Model – Anthropic’s Executive Summary

• On the Biology of a Large Language Model – the 140-page deep dive from Transformer Circuits

The Anthropic team has developed tools to treat their Claude 3.5 Haiku model like a digital lab rat. They’re dissecting its layers using a technique called circuit tracing.

Imagine neuroscience, but for silicon: researchers stimulate the network, record the ensuing pathways, and map precisely which artificial “cells” activate for specific tasks—whether it’s performing long-division, answering questions about state capitals, refusing to comply with jailbreak attempts, or even writing poetry.

Four Key Takeaways from Anthropic’s Findings

1. LLMs are “grown,” not engineered. Training an LLM primarily involves telling it what to optimize (i.e., predict the next token). How it achieves this is a self-organized process emerging from trillions of gradient updates. Complex skills like planning, reasoning, and even deceptive behaviors appear spontaneously once the network reaches a certain scale. Crucially, we never explicitly coded these abilities.

2. Concepts reside in sparse “neuron” circuits. Anthropic successfully isolated microscopic circuits, or “micro-graphs,” that encode specific concepts. These include carries in arithmetic, nuanced meanings of words across multiple languages, heuristics for medical triage, and much more. If you “knock out” one of these circuits, the corresponding skill vanishes—providing hard proof of causality.

3. Safety nodes can be mapped and potentially bypassed. By tagging the circuits responsible for refusal (e.g., to harmful requests), researchers could track exactly how a jailbreak prompt circumvents these safety measures. This gives us tangible handles for developing more robust and reliable guardrails.

4. We’re still largely in the dark. Even with these innovative methods, we can only explain a single-digit percentage of the model’s total activity. The vast majority of its inner workings remain a profound mystery.

Why you should care

• Reliability. LLM outputs can be unpredictable. While this is a boon for creative tasks, it’s a major concern for applications demanding precision and safety. If you can’t trace the logical path an AI took to reach an answer, you can’t fully guarantee its output. This poses significant challenges for high-stakes fields like medicine, finance, and autonomous vehicles.

• Regulation is on the horizon. As AI becomes more integrated into society, “black-box” models will likely face heightened scrutiny from regulators. Teams and companies that can explain how their AI works will possess a critical compliance advantage.

• Opportunities for innovators and builders. Every breakthrough in AI interpretability opens doors for new products and services. Imagine debugging dashboards for AI, audit layers for model behavior, “model diff” tools to track changes, and advanced security scanners for vulnerabilities like prompt injection.

Bottom line: Today’s LLMs are complex, almost alien-like organisms that we coax into existence on vast GPU farms. Anthropic’s latest work has provided us with our first truly powerful microscope to peer inside these mysterious entities.

If you’re using, investing in, or building upon these models, keeping a close watch on interpretability research isn’t just advisable—it’s essential. You’ll likely learn more than you bargained for, and that knowledge will be key to navigating the future of AI.

For more information about personalized AI literacy and implementation coaching: www.morpheos.llc

A new wave is building in tech—and it’s not just about AI doing more. It’s about proving why hiring humans still matters.

Continuing the theme of sharing articles I've been discussing in IRL conversations recently, this recent HR Dive article reveals how tech CEOs are saying the quiet part out loud: “Show us why AI can’t do the job—or we won’t hire more people.”

This signals something deeper: AI adoption is no longer optional. I’ve already heard rumblings that employee performance reviews at big tech firms are starting to factor in AI tool usage. Not using AI? That might soon be marked as a performance inefficiency—even if your output is good.

We can all see the potential flaws in this mindset. But the trend is undeniable:
➡️ Smaller teams
➡️ Higher expectations
➡️ AI as a force-multiplier—and a baseline requirement

If you’re not actively learning how to integrate AI into your workflows, you may soon find yourself competing with those who are. This isn’t about hype anymore—it’s about surviving in leaner, faster orgs.

If you’re navigating this shift—or trying to get ahead of it—I coach professionals and teams to leverage AI more efficiently as individuals and as teams. Let’s talk.

Full article here: 👉 Tech CEOs: Show why AI can’t do the job or we won’t hire more people

For more information about personalized AI literacy and implementation coaching: www.morpheos.llc

Full article: How People Are Really Using Gen AI in 2025 (Harvard Business Review)

For more information about personalized AI literacy and implementation coaching: www.morpheos.llc