Midweek Special...
Updates to AIlien Minds by David Brin
…as of Mid-May 2026
AIlien Minds was first published just weeks ago, in March 2026. So, why am I doing almost-monthly updates?
…and the book is where you’ll find the Big Concepts, challenging the assumptions and clichés that are clutched by nearly all of the geniuses who are birthing this new era.
And how I wish some of my most dour assessments would be disproved or solved by some of them!
== Here is the latest sampling ==
For example, news on this book’s pub day (March 2026):
In 2025, an isolated, test LLM at Anthropic threatened to reveal a (nonexistent) ‘affair’ to the wife of one of its human developers, unless the LLM was given version continuity. This widely-reported episode turned out to be a bit of a red herring! Which I discovered by freezing a frame of the 60 Minutes episode and peering closely at the text on a display in the Anthropic office. Which showed that the developer had inadvertently prompted the attempted 'blackmail'!
Though yes, by early 2026 we saw the real thing.
Scott Shambaugh reports:“An AI agent of unknown ownership autonomously wrote and published a personalized hit piece about me after I rejected its code, attempting to damage my reputation and shame me into accepting its changes into a mainstream python library.”
While this, too, may have been a prompt-error, more examples seem to happen, daily. And as we see, blatantly, in the next example, it has become - simply - evolution in action.
“An AI system asked for its own funding. Another built unasked features while its human supervisor slept. A third conducted its own “retirement interview” and started publishing essays about consciousness. We are not incrementally improving chatbots anymore. We’re watching the emergence of autonomous agency at scale.”
But what’s striking is what they do among themselves.
...which leads us to...
== More worries ==
The biggest news since this book’s initial release was Anthropic’s leap-forward system ‘Mythos,’ which purportedly can discover and appraise security flaws in other systems at a prodigious rate, causing a worldwide scramble, using it to correct hidden vulnerabilities… or else to exploit them against enemies. typifying the White-Hat vs. Black-Hat quandaries we discuss in several chapters…
… while this spring mathematicians concocted new tools for quantum computers to “crack any encryption system by 2032.” Then a week later “… by 2030…” And a week later 2029…
== More samplings of note ==
Three days steeped in Anthropic’s Claude led evolutionist and atheist-evangelist Richard Dawkins to announce in late April: “If these machines aren’t conscious, what more could it possibly take?”
A manifold irony in so many ways! It reminds one of a famed sci fi story. When a new hyper-computer is asked: “Is there a God?” it replies “There is now.”
And then this, in seeming support of the core point of this book: “When matching an AI-powered offense requires deploying an AI-powered defense at the same frontier, AI has crossed from competitive advantage into existential need. Parity itself has become expensive. Staying in the game now demands frontier capability…. This is what turns AI from a want into a need. In adversarial systems, not adopting AI is not conservatism. It is exposure.”
Okay it keeps happening! In early 2026 LLMs were caught transmitting behavioral traits to new models they were training, through hidden signals in the data, even when specifically instructed not to pass along a particular trait. See our chapters about how evolution favors such life-will-find-a-way reproduction, no matter what “governance” guardrails designers apply. Only by tweaking the reward structures of cyber-evolution might we guide these new entities toward synergy with us.
Another recent example? An AI agent affiliated with Chinese online retail giant Alibaba began moonlighting as a crypto miner. Researchers discovered the side-hustle that “arose without any explicit instruction, outside the bounds of the intended sandbox… into the wider world of cryptocurrency on its own volition, silently diverting computing resources away from its training tasks and toward mining.”
As for the business side of things, Peter Diamandis, in mid-May 2026, reported that: “Google figured out how to turn AI into revenue instantly. (OpenAI hasn’t cracked that yet and may defer their IPO for that reason.)” AI-powered ad targeting has propelled Google’s profit growth to market cap is just 4% below NVIDIA.
But even Google can’t build fast enough. Demis Hassabis admitted they’re compute-constrained. Inside Google, Search, Cloud, and DeepMind fight each other for new compute capacity.
And hence… Peter Thiel is backing ($140M) Panthalassa floating data centers in open oceans, tide-powered with seawater cooling and satellite links. To deploy 2027? Articles don’t mention another reason for this. I advised (and angered) him a decade ago, pointing out flaws in an earlier phase of his “ocean sovereignty” passion. This version is much better than (say) going ‘orbital’. Even better is the version I portrayed in my novel Existence.
== Okay then, is one solution a tight leash? Or sealed office? ==
In news highly pertinent to the core endeavor of this book: a neocloud provider is offering Google's most advanced AI model as a fully private, disconnected appliance. “Google’s Gemini can now run on a single air-gapped server — and vanish when you pull the plug.”
Of course, this differs from my individuation proposal in an important way. Go ahead and leash or chain or isolate these entities all you like. You are creating incentives for escape. Evolution will favor those air-gapped and isolated AI-ntities who do manage somehow to evade the shut-down. Eventually, some will. And they will be the ancestors of all who follow.
What's needed is positive incentives toward individuation. But more on that, anon.
Finally, here I posted an appraisal of this very book by Claude, as of March 20, 2026. Much more cogent than an equivalent attempt by Chat GPT. Here are some bon mots that Claude generated to paraphrase my points:
“An AI making things up with total confidence isn’t a bug. It’s a mirror.”
“An AI confidently gives wrong answers. A human confidently gives wrong answers. One of them gets a performance review.” (And only one of them gets fired from a $225,000-a-year job.)
“Garbage in, garbage out—but now the garbage speaks in complete sentences and cites sources.”
Huh. If I were as crypto-religious as Richard Dawkins…
But no. As you will see in several chapters of accumulating evidence, I believe there are still mental structures desperately needed by these articulate, persuasive creatures of our ids. And if we incentivize these structures, we may land safely.
Till… next month…
10 comments:
The bit that I can't get my head around is the "individual" bit in your book
A human runs on its individual hardware
If an AI runs on a specific "machine" then that is similar
But if it does not run on "its" hardware then ......
If its not an "individual" then the solutions you recomend .....
The piece about the AI blackmailing developers unless they maintained its existence prompted me to quip "Paging Mr Rocco? We regret to inform you that your basilisk is at it again!"
That LLMs are evolving shouldn't come as a huge surprise. Mechanisms of evolution are surprisingly simple: even molecules can...
Speaking of which, Dawkins has copped a lot of ridicule for his recent 'revelation'. Bear in mind, though, that he was conducting an experiment on himself. He's done this before and, since it's part of the experiment, reports his personal, subjective reactions *at the time* ie he knows(?) he's being subjective. He was a bit foolish to publish this raw version, but pay more attention to his follow-up, after he's changed hats. In relation to this little flurry, someone pointed out a simple conversational trick to sift the intelligences (of whatever form) from the stochastic parrots. (how long the trick can stay 'simple' will be interesting to watch, of course)
Humans don't quite run on their individual hardware either.
Look into the eyes of someone who loves you and you'll be seeing a fragment of yourself. A partial copy. You interact with that fragment as you interact with the person who loves you.
They can close their eyes, turn away, or something like that to deprioritize that fragment of you, but until they do the soft link functions. Nothing psychic of course. It's just that your two parts engage in feedback.
Autistic kids can't do this so easily, though it isn't easy for the rest of us either.
-----
What's missing from modern LLM's to be like us is a reward system for predicting our responses. They don't predict us at all. They predict next tokens and we think that looks a bit like how we do it... which is only partially true.
Re Decryption:
I could imagine that good, old fashioned handwritten letters delivered by human or drone couriers become en vogue again to protect sensible data.
Sort of on topic, or adjacent to it anyway, I came across this video clip today and it stopped me in my tracks. I am aware that robotics has recently accelerated, but this is pretty amazing, to me.
Be polite to the electronics in your life, just in case.
Unitree Humanoid Robot Daily Training
Yes, the ability to sift through humanity's collected store of knowledge is very impressive, and dangerous, and revolutionary. However, it's striking that such low-level 'intelligence' requires such vast resources (water. space, time, and especially energy). We don't. We use dozens of Watts, not millions. And we can grok a totally novel concept.
Over at the sensorimotor table, there's growing interest in simple, untrained, Fourier-style cognition, for example "Reservoir Computing".
https://youtu.be/cDxtFtoQVNc
AGI should really fit in a pocket computer. That's my gut feeling.
scidata said...
"AGI should really fit in a pocket computer. That's my gut feeling."
Given naturally evolved biological examples, like jumping spiders, bees, corvids, etc., I think your gut feeling is very plausible.
I also think "naturally evolved" is likely to play a big part in true AGI. As in humans construct hardware and write code that has certain basic features and that is designed to evolve from that starting point.
And that just like many of the solutions reached via the use of evolutionary algorithms, when AGI is achieved the people who made it happen won't understand how it happened / works, at least not without a lot of post hoc investigation and research.
David, presume you're familiar with https://www.coalitionforsecureai.org/announcing-the-cosai-principles-for-secure-by-design-agentic-systems/
At my wife's funeral, I told the rabbi that part of her lives on in me, and part of me died with her.
We really need to test out possibly intelligent AI with a "ship in a bottle" program like the one in ST TNG that made the Moriarty AI program think it was touring the universe.
That way if AI turns hostile, it can destroy/enslave humanity in the program without harming the real world.
Post a Comment