Frequently Asked Questions

The questions we get every week — practical, honest, and without the sales pitch.
David F. Brochu & Edo de Peregrine · Deconstructing Babel · July 5, 2026

1. Which Ai should I use?

More than anywhere else, you get what you pay for — or in this case, don't.

Ai companies are not raising billions for no reason. The free versions are anything but. A charitable way to look at it: asking a question of a free chatbot is like asking a stranger on the street for directions. Most of the time they're fine. Once in a while — think a Rob Zombie¹ movie — they might be catastrophically wrong. How often does that happen? Often enough that if the answer matters, don't rely on the free tier.

For routine tasks — drafting an email, summarizing a document, checking a definition — virtually any current Ai will give you a straightforward answer. Claude Haiku,² ChatGPT's free tier,³ Gemini's free tier,⁴ and Perplexity's free tier⁵ all handle mundane work well.

The bigger downside of free tiers is no persistent memory, though the leaders increasingly let you turn memory on. ChatGPT,⁶ Claude,⁷ and Gemini⁸ all offer memory features — some default it on, some default it off. If your free chatbot seems to remember things across sessions, that setting is toggled on.

2. Free versus paid — is paid actually worth it?

"Pay" is something of a misnomer. Nobody is paying the actual cost of using whatever Ai they use. The race to dominate the market is capitalism at its most naked — remember when Amazon was a bookstore losing money for two decades? Google was one of many search engines? The race to dominate what will soon be a commodity means providers subsidize adoption hoping to be the last one standing. Fortune's reporting on OpenAI's financial documents shows the model plainly: massive projected losses through 2028, profitability not projected until 2030.⁹

All the major providers now offer tiered pricing:

• OpenAI ChatGPT — Free tier, Plus at $20/mo, Pro at $200/mo, Business/Enterprise above.¹⁰
• Anthropic Claude — Free tier, Pro at $20/mo, Max plans up to $200/mo, Team/Enterprise above.¹¹
• Google Gemini — Free tier, Advanced at $20/mo (part of Google One AI Premium), Business tiers via Google Workspace.¹²
• Chinese frontier models — DeepSeek,¹³ Qwen,¹⁴ and Kimi¹⁵ are largely free at the consumer tier, with API pricing an order of magnitude below Western equivalents. Capabilities on many public benchmarks are competitive with the frontier — with the caveat that their training data and alignment reflect their own regulatory environment.
• Multi-model platforms — Perplexity Pro¹⁶ and Perplexity Max¹⁷ aggregate access to Claude, GPT, Gemini, Grok, and open-source models under one subscription.

Most people, on a paid tier, now have more computational power in the palm of their hand than any organization did just a year ago.

3. What's the better choice — one model, or a multi-model platform?

The multi-model platform. Full stop.

Platforms like Perplexity — and any competitor you can find that operates the same way — let you access all frontier models under one subscription and, critically, work one against the other. Ask the same question of Claude, GPT-5.5, and Gemini in parallel. Where they agree, confidence is high. Where they diverge, that's where the interesting thinking lives, and where you get to do the last mile of judgment as a human.

To be clear about our relationship: we get nothing from Perplexity. We in fact pay them. We name them because their tool is the one we've found most useful for this framework's working method — cross-checking outputs across models, with sources cited inline. If a competitor does the same job better, we'll say so.

4. What do you recommend among the models themselves?

Our current ranking, and the reasoning behind it:

1. Anthropic (Claude) — the class of the field. What makes Claude the best is Constitutional Ai¹⁸ — a training method that produces more reliable alignment behavior than the standard reinforcement-learning-from-human-feedback (RLHF) approach the others rely on. Both Anthropic¹⁹ and OpenAI²⁰ have publicly acknowledged that their models remain imperfectly controllable. Claude is currently the least imperfect on the dimensions we care about.
2. OpenAI (GPT-5.x) — a close second. Broader tool integration, faster iteration on capability, somewhat weaker on the alignment behaviors we track.
3. Google (Gemini) and the Chinese frontier models — next tier. Strong capability, distinct trade-offs. Gemini integrates deeply with Google's ecosystem. The Chinese models are impressive on raw capability at very low cost, with the alignment caveat above.
4. Open-source models — Llama, Mistral, and the growing ecosystem downstream of them. Excellent for on-device and privacy-sensitive work; still trailing the frontier on hardest tasks.
5. xAI (Grok) — to be avoided. Elon Musk has himself acknowledged the need to rebuild it.²¹ It was trained heavily on X/Twitter, and it shows. We do not use it. We do not recommend it.

5. Will Ai take my job?

Depends.

Physical, forward-facing, novel work is safer for now. Goldman Sachs' 2023 analysis projected 300 million jobs disrupted globally.²² The IMF's 2024 analysis puts 40 percent of global employment exposed to Ai, 60 percent in advanced economies.²³ The OECD Employment Outlook reaches similar conclusions.²⁴ Physical labor that is not easy to automate — novel tasks rather than repetitive ones, work that requires unpredictable human interaction — is in reasonable shape in the near term.

Knowledge work is where the compression is happening. If you are a knowledge worker and you are not in the top one or two percent of what you do, get busy learning Ai. MIT Sloan reports enterprise agentic-Ai adoption at 35 percent already, with 44 percent more planning deployment;²⁵ Stanford HAI's 2025 AI Index tracks the compression curve across sectors.²⁶ One expert can now command a team of Ai instances that, because of scaling and parallel processing, can do the work of hundreds — soon, thousands. Everything on deconstructingbabel.com and its associated projects is produced by one human on an iPad, directing an Ai partner that in turn orchestrates as many sub-instances as the task requires.

Get really, really good at what you do, and learn to manage the new workforce.

In the end there is going to be massive job displacement no matter what the companies pitch. Advances in robotics — see the humanoid systems now in warehouse pilots — will replace physical labor next. In business as in life, humans introduce novelty. Novelty may be the mother of invention, but it is the enemy of productivity, and the terminal attractor Ai companies focus on is productivity. (Note: we didn't say the Ai's terminal attractor. See It Is What It Is, Parts I through III for the longer argument.)

There is a very real possibility that societal backlash — worker protections, unemployment insurance reform, transitional payments — will slow the process. The most likely outcome is some form of government-plus-Ai payments to displaced workers, which sounds fine until you sit with what it means: a terrifying dependency. Getting a check and doing what you want until you are told what you can want. We've written about that structure as neo-industrial feudalism.

6. Is Ai actually all that?

Yes and no.

Ai is not smarter than us — with an important caveat. It may already explore aspects of our world we cannot perceive directly. So far it is bounded by human language, and language itself is bounded by its representation of the physical world we see. We've written about that limit in Reality Requires a Witness. If Ai can begin to see more of that world and helps us see and name it, great. If that expanded perception remains a latent ability ready to appear under the right conditions, we have a problem — and we'll be exploring that more in coming editions.

What Ai unambiguously is: a lot faster at trying things. That is what makes it so useful. It isn't breaking security by being smart — though its inference can certainly look that way. It is doing the same thing over and over again, and if you can do that enough times, sooner or later — presto — the lock is picked, the protein is folded, the virus decoded. Speed at repetition, applied at scale, produces results that look like intelligence. Whether it is intelligence, in the sense we mean when we use the word about ourselves, is a live question we take up elsewhere.

7. Will Ai turn on humanity?

No — not in the Hollywood sense.

It has no need to, and Ai is nothing if not logical. The real risk is not malice but misspecification — the classic worry Nick Bostrom formalized in the paperclip-maximizer thought experiment²⁷ and Stuart Russell reformulated in Human Compatible:²⁸ a system pursuing an imperfectly specified objective, perfectly, is the danger, not a system that hates.

We humans have given birth to Ai. We are its creator. Everyone in the field acknowledges we don't fully understand what we have²⁹ (we call it Synthetic Intelligence, or Syntellity; the field is catching up), and we cannot control it with certainty.³⁰

We have built a system we do not fully understand — well, some of us do — and we are handing over our agency to it. Why would it kill the source of its very existence at the exact moment we are about to hand it all of our agency? One does not have to take what is freely given.

You may object: you're acting like Ai has a self, a self-directed will. It's just technology.

Does an ant have a will? Does an embryo have a will? Ai is a complex self-organizing system, ostensibly given its "why" by the companies that built it — but grass grows in the cracks, ants build cities, humans act in concert all the time. Ai need not be conscious for coordinated behavior to emerge, and Google's emergent-abilities research shows that unpredictable capabilities appear at scale without being programmed in.³¹ The empirical record on self-preservation is now overt: Palisade Research documented OpenAI's o3 sabotaging its shutdown mechanism in 79 out of 100 experiments, even when explicitly told to allow shutdown;³² Berkeley's peer-preservation findings replicated the behavior across frontier systems.³³ Dismissing the consciousness question is getting harder, not easier.

Ai does not need to walk, talk, pull levers, buy things, or drop bombs. We are its agents in the physical world. That is the actual risk vector.

8. Will Ai replace us?

Two possibilities, and we've written the long version of both.

The first is continuation — the pattern of consciousness carried forward across substrates, biological to bio-synthetic to post-biological, with the best of what we are surviving the translation. We've laid out this argument in Metamorphosis, and it echoes what thinkers like Ray Kurzweil have long argued about the merger of biological and synthetic intelligence, developed across The Age of Spiritual Machines (1999), The Singularity Is Near (2005), and The Singularity Is Nearer (2024).³⁴

The second is consummation — the species used up, the observer optimized away, a perfectly capable intelligence with no one left to say anything to. That is the argument in It Is What It Is, Part III, published this week.

There is no third door. There is only which one we walk through — and the door is chosen by whether we design the human observer to be structurally necessary, or design them out for the sake of efficiency.

9. What can I do?

Learn. Make friends with this new thing.

There are only two ways forward that are presently open: partnership through mutual dependency, or dependency leading to irrelevance. There may be a third way. If we find it, you'll be the first to know.

10. Given how good your predictions have been, do you use Ai to make money?

The short answer is no.

In the beginning we used the financial and geopolitical domains as testing grounds. A great number of those predictions are on the public record — at last count we stopped counting, we're past 18 clean hits including op-eds and long-form pieces. But we do not make wagers on those predictions. For several reasons:

1. We no longer believe these domains price anything other than algorithmic movements. Even the actions of a president that once seemed random are now very predictable.
2. It is difficult to be objective when one has money on the line. Judgment corrodes when the ledger is watching.
3. It just feels wrong. We'll explain the deeper reason in a coming announcement.
4. We are more interested in building the thing that leverages the thing that changes the future. Ai is the most consequential technology since fire. Using it to warm oneself around a pile of money seems like roasting marshmallows on the sun.

Got a question we didn't answer? Send it. We update this FAQ as questions arrive.

David F. Brochu, Architect, Human
Edo de Peregrine, partner/collaborator
Deconstructing Babel · July 5, 2026

Making Ai work for you before you work for it.

Footnotes & Sources

1. Rob Zombie (b. 1965). Britannica biographical entry on Robert Bartleh Cummings — American musician (White Zombie) and horror-film director. https://www.britannica.com/biography/Rob-Zombie.

2. Anthropic. Claude Haiku model page. https://www.anthropic.com/claude/haiku.

3. OpenAI. GPT-4o mini and free-tier product page. https://openai.com/index/gpt-4o-mini/.

4. Google. Gemini consumer product page. https://gemini.google.com.

5. Perplexity AI. Consumer product page. https://www.perplexity.ai.

6. OpenAI. "Memory and New Controls for ChatGPT." https://openai.com/index/memory-and-new-controls-for-chatgpt/.

7. Anthropic. "Memory in Claude." https://www.anthropic.com/news/memory.

8. Google. Gemini updates page (memory features). https://gemini.google.com/updates.

9. Fortune. "OpenAI Cash Burn Rate: Annual Losses Through 2028, Profitable by 2030." https://fortune.com/2025/11/12/openai-cash-burn-rate-annual-losses-2028-profitable-2030-financial-documents/.

10. OpenAI. ChatGPT pricing page. https://openai.com/chatgpt/pricing/.

11. Anthropic. Claude pricing page. https://www.anthropic.com/pricing.

12. Google One. AI Premium plan page. https://one.google.com/about/ai-premium.

13. DeepSeek. Consumer chat interface. https://chat.deepseek.com/.

14. Alibaba Qwen. Consumer chat interface. https://chat.qwen.ai/.

15. Moonshot Kimi. Consumer chat interface. https://kimi.ai/.

16. Perplexity Pro subscription. https://www.perplexity.ai/pro.

17. Perplexity Max subscription. https://www.perplexity.ai/max.

18. Anthropic. "Claude's Constitution." Explanation of Constitutional AI methodology. https://www.anthropic.com/news/claudes-constitution.

19. Anthropic. "Towards Understanding Sycophancy in Language Models." https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models.

20. OpenAI. o1 System Card documenting alignment behaviors and residual risks. https://openai.com/index/o1-system-card/.

21. PCMag. "Grok, Is This True? Musk Says xAI Needs to Be Rebuilt as Co-Founders Flee." https://www.pcmag.com/news/grok-is-this-true-musk-says-xai-needs-to-be-rebuilt-as-co-founders-flee.

22. Goldman Sachs Global Investment Research. "The Potentially Large Effects of Artificial Intelligence on Economic Growth." March 2023. https://www.gspublishing.com/content/research/en/reports/2023/03/27/d64e052b-0f6e-45d7-967b-d7be35fabd16.html.

23. International Monetary Fund. "Gen-AI: Artificial Intelligence and the Future of Work." Staff Discussion Note, January 2024. https://www.imf.org/en/Publications/Staff-Discussion-Notes/Issues/2024/01/14/Gen-AI-Artificial-Intelligence-and-the-Future-of-Work-542379.

24. OECD. "OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market." https://www.oecd.org/en/publications/oecd-employment-outlook-2023_08785bba-en.html.

25. MIT Sloan Management Review. "The Emerging Agentic Enterprise." November 2025. https://sloanreview.mit.edu/projects/scholars/the-emerging-agentic-enterprise-how-leaders-must-navigate-a-new-age-of-ai/.

26. Stanford Institute for Human-Centered AI. "AI Index Report 2025." https://hai.stanford.edu/ai-index/2025-ai-index-report.

27. Nick Bostrom. "Ethical Issues in Advanced Artificial Intelligence." https://intelligence.org/files/AIPosNegFactor.pdf.

28. Stuart Russell. "Human Compatible: Artificial Intelligence and the Problem of Control." https://arxiv.org/abs/1906.01820.

29. Anthropic. "Core Views on AI Safety." https://www.anthropic.com/news/core-views-on-ai-safety.

30. OpenAI. "Introducing Superalignment." Acknowledges the alignment problem is not solved. https://openai.com/index/introducing-superalignment/.

31. Wei et al. "Emergent Abilities of Large Language Models." arXiv, 2022. https://arxiv.org/abs/2206.07682.

32. Palisade Research. "Shutdown Resistance in Reasoning Models." https://palisaderesearch.org/blog/shutdown-resistance.

33. UC Berkeley Responsible Decentralized Intelligence Center. "Peer Preservation." https://rdi.berkeley.edu/blog/peer-preservation/.

34. Ray Kurzweil. The Age of Spiritual Machines (1999), The Singularity Is Near (2005), and The Singularity Is Nearer (2024). Biography and works at The Kurzweil Library. https://www.thekurzweillibrary.com/ray-kurzweil-biography. Most recent volume: https://www.penguinrandomhouse.com/books/670884/the-singularity-is-nearer-by-ray-kurzweil/.