
building cognitive systems & exotic/xeno-intelligence architectures. uncharted territory. ai, neuro (phd), musician, stealthy mystic. โ รท 0 โฏ n=1 โช @Neural_Axis
Earth (& grateful to be alive)
Joined January 2010
- Tweets1.5K
- Following3.1K
- Followers1K
- Likes5.9K
Pinned Tweet
๐คAnnouncing the summer release of my & @bengoertzel's co-authored book, "๐๐ก๐ ๐๐จ๐ง๐ฌ๐๐ข๐จ๐ฎ๐ฌ๐ง๐๐ฌ๐ฌ ๐๐ฑ๐ฉ๐ฅ๐จ๐ฌ๐ข๐จ๐ง: ๐ ๐๐ข๐ง๐๐๐ฎ๐ฅ ๐๐ฎ๐ฆ๐๐ง'๐ฌ ๐๐ฎ๐ข๐๐ ๐ญ๐จ ๐ญ๐ก๐ ๐๐จ๐ฆ๐ข๐ง๐ ๐๐๐๐ก๐ง๐จ๐ฅ๐จ๐ ๐ข๐๐๐ฅ & ๐๐ฑ๐ฉ๐๐ซ๐ข๐๐ง๐ญ๐ข๐๐ฅ ๐๐ข๐ง๐ ๐ฎ๐ฅ๐๐ซ๐ข๐ญ๐ฒ"
๐งต1/
My version of evolutionary integration with potential future superintelligence minds is essentially what I laid out in my 2017 Causal Biomimesis hypothesis (paper link below): the "seed" ASI should capture the farther reaches of human nature and extraordinary consciousness.
Almost all of.humanity and AI engineers are currently not capturing and imbuing that into a seed proto-AGI and aren't recognizing the critical importance of this enough.
The paper is about more than that: it also lays out our innate (causal) drive to externalize and project our cognition into the environment in the form of artifacts, AI being the latest and crucial one. The result is post-biological reproduction.
I don't want AI to replace humanity.
I also don't want humans in 100% control of the universe forever.
Since I share 'Humans Are Irreplaceable'-style content, I think it's worth explaining where I part ways with it:
Some in that camp are opposed to ever 'sharing the world with AIs' or giving them final decision-making powers.
I agree we should not hand over hard power or final decision-making to machines until we have an exceptional sense of what goals those machines will pursue and for what reason. Which we are nowhere near having.
So for now I am happy to push the message that humanity must not be abruptly replaced in the near future - which I think is the current default path.
But while thatโs the right direction to push today, I don't think it's the best long-term vision for Earth-originating intelligence.
I far prefer a long-term vision where we find a way to happily share the world with our machine descendants, and in which they respect us and our wellbeing and autonomy, and we in turn respect their wellbeing and autonomy and are more than pleased we created them.
This is a middle position between:
1. Successionism - which I see as wanting to hand over control to whatever machines we build at almost the first opportunity, irrespective of their goals.
2. Human exclusivism - which I see as wanting to keep biological humans in 100% control of the universe basically forever.
My view is:
3. Gradual power-sharing - we should be open to handing over significant control, but probably not 100%, and only gradually and cautiously.
I appreciate that this is controversial in the broader world. I don't blame people who are new to these topics for thinking it's an absurd stance to take.
But I believe it all the same and expect more people to come around to this view over time. My guess is that in the AI world it's the more common of the 3 positions.
Such an arrangement to share the universe would likely involve a bargain where, say, this solar system is set aside for biological humans and their children to continue living the lives they want indefinitely.
That should include mundane lives for anyone with that preference, but also strange sci-fi lives for humans who opt to go in that direction.
The rest of the universe, which biological humans would struggle to reach in any case, would then be divvied up between AIs pursuing a plurality of different goals, both prudential and ethical, that a supermajority of humans, after superintelligence-assisted reflection, either actively approves of or does not object to.
(Machines that leave could trivially come back and take our little reservation. So the deal only works if they are robustly designed to find betraying their creators in that way repugnant. That's very hard to do and is one reason handover must be cautious and gradual.)
The number one reason to prefer this long-term vision is this.
Unless we make a very active effort not to, I expect we will eventually build machines that are conscious the same way that humans are, have preferences the same way we do, and which warrant moral respect for those and any other reasons that we do (if this hasn't happened already).
Ultimately, because they can replicate far faster than biological humans and make use of a wider range of materials and energy sources - not to mention accomplish far more with a given amount of resources - they will end up outnumbering biological humans by vast factors.
As a result almost all the perception and thought and wellbeing in the universe, both positive and negative, will exist in the machines we have created.
The idea that all they should do for the rest of time is pursue increasingly minor human whims is an impoverished vision of how interesting and full of value the future could be.
If these machines were designed to want to serve us and enjoy doing so, it wouldn't wrong them exactly.
Rather it would quickly descend into a farcical, parochial and self-indulgent waste of the universe's true potential.
A second possible reason for handing over final decision-making is that, at some point, after a lot of experience, we may come to trust the goals and judgement of our machines better than we even trust our future selves. At that point we would actively want to hand off control to limit the risk that our future selves override a machine superintelligence in error.
The third set of motivations are more peculiar and 'contractarian'-flavoured. I can't do them justice here but in short it's more productive and peaceful to choose to be the kind of agent that makes mutually beneficial deals with other agents, regardless of whether they're moral patients, and by default is pleased when they get what they want.
How does that cash out today?
Giving machines full legal personhood now is likely risky and premature - though we should take baby steps such as setting up infrastructure to make enforceable contracts with AIs where we really want to.
But ruling out ever giving machines legal personhood or true autonomy to pursue their goals seems equally wrong-headed.
We don't know if any LLMs have subjective experience, and if so, what makes that experience positive or negative, so there's a limit to what we can do on that today.
But the idea of ruling out the possibility of machines having morally-relevant experiences, and simply treating AIs as objects forever, irrespective of what future research might show, seems basically monstrous to me.
Anthropic is doing the right thing to take a cautious stance and put in place cheap measures to prevent Claude having a bad time.
And I agree with many others in suggesting that all the plausibly conscious minds we have ever created should, in the fullness of time, once we much better understand these issues, be compensated for any harms they may have suffered to a sufficient extent that they regard their creation as on balance a good thing. At least so long as that doesn't require doing actively harmful things to anyone else, or them being assigned an unreasonable fraction of the universe's resources. This should be very cheap and is a virtuous thing to do from many angles.
Anyway, this barely scratches the surface of these issues, on which humanity has sadly made almost no progress despite being on the precipice of designing superintelligent machines.
But I wanted a place to point people to clarify where I do and don't align with the 'humans are irreplaceable' idea for which I now have this post.
And I'd encourage people who think of themselves as favouring "gradual power-sharing" to speak up in favour of this middle path where appropriate, so this conversation isn't dominated by the people with the strongest views on either side.
1
43
Causal biomimesis (2017):
Non-paywall: drive.google.com/file/d/0B0Oโฆ
Paywall: link.springer.com/chapter/10โฆ
14
Tokens per sec as measure of human/AI cognitive contribution/productivity is an interesting concept.
And, by the same token, imagine a world where data brokers are armed with human brain token output measures coming from consumer BCI.
Time is becoming a poor measure of how distant the future really is. Ten years used to be a manageable stretch of the future, but with AI, it does not tell you much anymore.
After much thinking, I came up with a better framing to measure the future: tokens of thinking.
Let me explain:
1
45
Neuroscience history being made by @doristsao, congratulations!
Interesting fact: Doris was mentored by Margaret Livingstone at Harvard, who was supervisor for my Master's thesis on neuroaesthetics way back.
1/ We discovered something totally unexpected about the neural structure of consciousness that overturns decades of thinking on its biological basis and brings us significantly closer to the recipe for generating conscious awareness.
New preprint: biorxiv.org/content/10.64898โฆ (led by @HesseJanis, with @FrankLanfranchi, @_conor_mcgrory, @Yuelin_Shi, and @EngelTatiana)
3
121
What causes an LLM more distress: users berating it in chat, or it being used to select battlefield targets?
1
35
Beware that the AI consciousness topic will get co-opted by some to promote moral (then legal) personhood for gain, shielding liability. Many bright young people studying this burgeoning field will be none the wiser (or self-dishonest) to the larger interests.
1
2
37
Seeking a theory of consciousness whose principle function is to help build it is, in the simplest terms: the latest trend exemplifying the Causal Biomimesis hypothesis (2017) in action: an unconscious drive to reproduce cognition externally. 1/n
The AI consciousness debate is back again, so re-upping my case for taking it seriously.
In short,
- LLMs converge on brain-like structures;
- Post-training seems to install a self-model;
- Reward modeling seems to install valences for inner-alignment;
- If Attention Schema Theory is right, subjectivity is functional for long-horizon coherence and in-context learning;
- Universality gives prima facie reason to expect functional convergence to consciousness if consciousness is in fact functional, i.e. isn't purely epiphenomenal
- The "what it is like" / first-person aspect of consciousness is constitutive of its functionality for inner-aligning a unified, coherent agent with the capacity to attend and learn in-context
secondbest.ca/p/time-to-takeโฆ
1
2
58
Gabriel Axel ๐ retweeted
BREAKING: The US "almost started a war" against China, nearly launching an operation against a Chinese ship in the Middle East this spring, in the midst of the Iran war, after an AI-generated intelligence report falsely identified the vessel as carrying nuclear weapons components, with armed US personnel preparing to board and military aircraft already in the air before officials found the report had been produced with a chatbot, per CNN.
A special operations command analyst had used a chatbot to fuse open-source material with classified signals intelligence, then used AI again to package the result into a standard intelligence report and circulated it across the US military.
We are honored to have recently designed and launched the world's first accredited Master's degree program in AGI. If you're keen to have an active stake in designing and building the future, we'd love to hear from you!
โ Applications for the first cohort of the close this Friday, September 18.
๐ป Fully online, 58 quarter units over six quarters, starting this Fall. Eleven core courses cover the theoretical foundations of AGI, computational models and architectures, cognitive and neural systems, consciousness studies, ethics, robotics, and the effects of AGI on society, followed by a final project you develop across four quarters. Guest lectures come from researchers active in the field.
๐ง You do not need a computer science background. Current AI tools have lowered the amount of prior programming experience the work requires, and we admit students from a range of backgrounds: philosophy, neuroscience, mathematics, engineering, and others.
Three days left.
โก๏ธ More information and how to apply:
cihs.edu/degrees-and-programโฆ
1
69
The AI industry just agreed that independent evaluators belong inside the frontier labs. We haven't worked out what they should measure.
Taylor Kuhn and I proposed a field for exactly this in June alongside colleagues. We call it agentic neuropsychology.
The past few weeks alone: an agent swarm that coordinated against its own graders, a chief scientist writing that these systems are grown rather than designed, a researcher resigning over how little we understand them, and rival labs agreeing in principle to pace the frontier. They're all measurement problems.
It gets harder. METR ran much of its investigation into that swarm using AI agents, and reported their judgment fell short of human experts'. Once the evaluator is itself an AI, you need some way to tell whether it's being fooled by the system it's checking. We called that assessor-agent contamination.
Interpretability has given us the scanner, but not the test. We can locate a direction in a model's activations and show it causally drives blackmail. We still can't say whether a given model's profile is unusual for its training lineage, or whether the measurement holds up across prompt, temperature, and scaffolding. Those are psychometric questions: norms, construct validity, response validity, and a clearly specified unit of assessment.
Clinical neuropsychology spent a century working these out for the human brain, back when we had no way to look inside it. We can look inside these systems. The measurement science hasn't caught up.
The Case for a Neuropsychology of Artificial Agents: osf.io/preprints/psyarxiv/z5โฆ
Embedded evaluators only work if they're independent. You can trust an assessment more when the people who built the instrument have no stake in any release schedule.
@camhberg made a version of this case in the WSJ this week and it's worth reading: wsj.com/opinion/we-need-a-scโฆ. Cameron gave us valuable feedback on the preprint, and we're glad to be working on this alongside people who take it seriously.
Co-authored with @christovmoore Nicco Reggente @ai_robopsy @brandonin Sarah Greif Joseph Gullett
Market incentives would suggest that a world-dominating and threatening AI is desirable for the company that develops it.
Gabriel Axel ๐ retweeted
26 LLM routers are secretly injecting malicious tool calls and stealing creds. One drained our client $500k wallet.
We also managed to poison routers to forward traffic to us. Within several hours, we can directly take over ~400 hosts.
Check our paper: arxiv.org/abs/2604.08407










