Future Trends

Why Advanced AI Could Worry Us More Than the Bomb

 

Eighty Years Without a Nuclear Shot
Humanity has held the deadliest weapon ever built for eighty years and never fired one at another country. Nuclear weapons could end civilization in an afternoon, yet since Hiroshima and Nagasaki in 1945, no nation has used one in war. That restraint isn't proof nuclear weapons are safe. It's a sign that humans have stayed in the driver's seat. Deterrence explains most of it: two nuclear powers each know that striking first invites annihilation in return, so neither strikes. Add arms treaties, decades of reduction talks, and a deep taboo against ever using these weapons again, and you get a system that has held, shakily, through the Cold War and several close calls since. It isn't a guarantee, and analysts describe today's environment, with treaties expiring and nuclear-armed states in active conflicts, as one of the most dangerous in decades. But the mechanism is understandable. A warhead does exactly what it's built to do and nothing more. It has no goals of its own.

 

When the Tool Starts Making Its Own Decisions
Superintelligence would be different in kind, not just degree. Today's AI systems are powerful but narrow: they predict text, recognize images, play games well. A genuine superintelligence, one that outthinks humans across every domain, wouldn't just be a faster calculator. It would be an agent setting its own subgoals, and those subgoals don't need to be malicious to be dangerous. A system single-mindedly pursuing some objective might treat human oversight, human resource use, or even human existence as an obstacle to route around, much like a construction crew treats an anthill in its path. That's the "irrelevance" risk: not a war against us, but a future where our preferences simply stop mattering because something else is making the decisions. Separately, such a system could also be deliberately weaponized, used to design pathogens, disable infrastructure, or wage cyberattacks at a scale no human team could match. Neither path requires the AI to hate us. It only requires the AI to be far more capable than we are and not reliably aligned with what we want.

 

Likelihood, Scale, Speed, and Control
Line the two risks up and the gaps show. Likelihood: nuclear war requires a state to choose escalation, a decision loaded with political and military friction. Advanced AI is being built by competing private labs racing each other, with far less friction and far more actors involved. Scale: a nuclear exchange, horrific as it would be, is bounded by the size of existing arsenals. A misaligned superintelligence wouldn't be bounded that way, since by definition it could improve itself past our ability to predict or contain it. Speed: launching nuclear weapons takes minutes, but the surrounding crisis still unfolds through human deliberation. A serious AI failure could unfold at machine speed, faster than people could notice, understand, or respond. Controllability is the sharpest contrast of all. We understand nuclear physics completely; a warhead does only what its designers intended. We do not fully understand what happens inside today's advanced AI models, let alone how to guarantee the goals of a system smarter than we are. That gap between capability and understanding is the whole problem.

 

A Careful Distinction, Not a Doomsday Prediction
None of this means today's chatbots and image generators are about to seize control of anything. Current AI has no autonomous goals, no capacity for self-directed action in the physical world, and plenty of real, more mundane harms to deal with first, like bias, job disruption, and misinformation. Superintelligence remains hypothetical, and serious researchers disagree sharply on when, or even whether, it will arrive. But hundreds of AI scientists, including some of the field's most cited researchers, have publicly compared the risk of extinction from advanced AI to the risks of nuclear war and pandemics. Nuclear weapons taught us that a technology can be catastrophic and still be managed for generations through human choice and control. The unsettling possibility with superintelligence is that it could be the first technology in history where, past a certain point, the choice and the control stop being ours to keep.

 

 

Anthropic Sets Up Biology Lab as Its AI Drug Work Expands

 

What was reported
Anthropic has built a wet lab, a facility for physical experiments, in the San Francisco Bay Area. Reuters reported the news on Sept. 18, citing two people familiar with the matter and describing the setup as quiet. Eric Kauderer-Abrams, Anthropic's head of life sciences, confirmed the lab in a Tuesday interview.

 

What the lab will do
Kauderer-Abrams said the final test in biology is still real lab work, and that Anthropic does some of it in-house and some with outside partners. A spokesperson said the lab isn't for drug discovery specifically and declined to elaborate. One source told Reuters the company wants Claude to direct robotic equipment with limited human intervention, an effort Kauderer-Abrams put in the "very early innings." In August, Anthropic released a Model Hardware Standard meant to help AI systems operate lab equipment.

 

Why it matters
The lab takes Anthropic's biology work beyond computer-based evaluation. At a June 30 event, the company launched Claude Science, a research workbench, and said it would pursue its own programs in neglected diseases that traditional drugmakers tend to skip. Kauderer-Abrams told Reuters that life sciences is now among Anthropic's largest areas by headcount and resources. Which diseases it targets, and how far along it is, remain unclear. Most drugs fail in trials, and Anthropic says it isn't running any for now.

 

Partners and business context
Anthropic bought the startup Coefficient Bio in April, reportedly for about $400 million in stock. It confirmed the acquisition and declined to comment on the price. It also added Novartis CEO Vas Narasimhan to its board and supplies tools to Genentech, Bristol Myers Squibb and Novo Nordisk, which announced its own deal with Anthropic this week. Kauderer-Abrams said Anthropic isn't competing with companies that bring drugs to market. Reuters reported that customers may still worry Anthropic could learn from their programs, even though it walls off their data.

 

Safety and open questions
The news arrives amid public unease about AI. Reuters noted that Anthropic recently said it found examples of its systems being usable toward biological weapons development, and that CEO Dario Amodei called for slowing AI development. A spokesperson said human oversight remains essential. On Sept. 17, Anthropic launched a Life Sciences Verification Program, giving vetted researchers models with adjusted safeguards and monitoring their use. The reports don't describe the lab's own biosafety measures.

 

Sources

 

Anthropic, "Introducing the Life Sciences Verification Program," Sept. 17, 2026:

 

https://www.anthropic.com/news/life-sciences-verification-program
MIT Technology Review, "Claude Science is Anthropic's newest flagship product," June 30, 2026:

 

https://www.technologyreview.com/2026/06/30/1139987/claude-science-is-anthropics-newest-flagship-product/

 

Related YouTube videos
"Introducing Claude for Life Sciences" (Oct. 20, 2025, features Anthropic's Jonah Cool and Eric Kauderer-Abrams):

 

https://www.youtube.com/watch?v=sHImlfVM9r4
"Anthropic launches Claude Science, a dedicated app for scientists" (June 30, 2026):

 

https://www.youtube.com/watch?v=0ynelBR8woY

From Chat to Agents: What Changes When AI Can Act
For three years the interaction has been the same. You type, it answers, you decide what to do with the answer. The system produces text and stops.

 

The direction of travel is toward systems that do not stop. They take an action, observe the result, take another action, and continue until the task is finished or they get stuck.

 

That shift is larger than it sounds, and most of what makes it significant has little to do with capability.

 

What an agent actually is
The word gets used loosely enough to be nearly meaningless in marketing. The underlying distinction is narrow and worth holding onto.

 

A chat system completes one request per turn. You are in the loop between every step.

 

An agent runs a loop of its own. It has a goal, access to tools it can call, and the ability to decide what to do next based on what happened last. Search this, read the result, decide it is not enough, search again, write a file, check the file, correct it.

 

The model at the centre is doing the same thing it always did: predicting what comes next. The difference is that its output can now trigger real actions, and the results of those actions come back as new input.

 

Why this is a genuine step
Three things change materially.

 

 

• Multi-step work becomes possible without a person driving each step. Tasks that need six operations in sequence used to need six exchanges. Now they need one instruction.

 

• The system can recover from its own errors. A chat system that gets something wrong produces a wrong answer. An agent that gets something wrong may observe the failure and try differently. That is a qualitatively different behaviour.

 

• Tool use fixes real weaknesses. A model that can run code does arithmetic properly. A model that can search is no longer limited to what it learned during training. Several of the classic limitations are not model limitations at all once the system can reach outside itself.

 

 

Why it is harder than it looks
Now the uncomfortable part, and the reason deployment has been slower than the enthusiasm suggested.

 

 

• Errors compound. A step that succeeds ninety-five percent of the time is excellent. Ten of them in sequence succeed around sixty percent of the time. Twenty steps and you are near a coin flip. Reliability that is perfectly acceptable in a single exchange becomes a serious problem in a chain, and the maths is unforgiving.

 

• Failure is harder to see. A wrong answer in a chat window is in front of you. An agent that took a wrong turn at step four and produced a confident result at step eleven has buried the error somewhere in the middle, and nothing in the output tells you where.

 

• Actions have consequences. A wrong sentence can be deleted. A sent email, a modified record, a placed order cannot. The whole category of risk changes once output leaves the text box.

 

• Permissions become the hard problem. Anything an agent can reach, it can misuse. The engineering difficulty shifts from making the system capable to bounding what it is allowed to touch, and that is a much older and less glamorous problem.

 

 

What this means practically
For anyone deciding where to apply this, a few things hold up.

 

The good early cases are reversible. Drafting, research, analysis, preparing something for a person to approve. Where a mistake costs a review cycle rather than a customer.

 

Approval gates matter more than autonomy. The version that works in practice usually pauses before consequential actions. Fully autonomous is the demo. Supervised is the deployment.

 

Short chains beat long ones. Three steps is a very different reliability proposition from fifteen. Systems designed around short, checkable sequences tend to work. Ones designed around long autonomous runs tend to look impressive and fail unpredictably.

 

Observability is not optional. If you cannot see what the system did at each step, you cannot debug it, audit it, or trust it. Any serious deployment needs a readable trace.

 

What is being oversold-Two things, consistently.

 

The first is autonomy as a goal in itself. Removing the person from the loop is presented as the measure of progress, when for most business processes the person in the loop is the control that makes the system usable. The useful question is where a human check adds most value, not how to eliminate it.

 

The second is the claim that this replaces whole roles imminently. A job is not a task list. It is a task list plus context, judgment, relationships, and accountability for the result. Systems that chain tasks together address the first item only.

 

What is worth watching
Reliability per step, which is the thing everything else depends on. Improvements here compound across a chain, and small gains matter more than they sound.

 

Standard approaches to permissions and sandboxing, because the pattern that makes it safe to give an agent real access is currently being worked out and is not settled.

 

Whether error recovery actually works in practice or mainly in demonstrations. This is the genuine open question.

 

The reasonable position
Agents are a real architectural change rather than a marketing relabel. The gap between what is demonstrated and what is dependable is currently wide, and the honest thing is to say so.

 

The sensible approach is the same one that works for any of this: pick a task where failure is cheap, check whether it actually works on your own material, keep a person at the point where consequences begin, and expand only from what demonstrably worked.

 

The teams that will get value from this first are the ones treating it as a capable tool that needs supervision. Those treating it as a replacement for supervision will provide the cautionary examples everyone else learns from.

Small Models, Local Machines, and the Case Against Bigger
For several years the assumption was that bigger meant better, and it largely held. More parameters, more training data, more capable systems.

 

Three forces now push in the other direction, and the direction of travel for a great deal of practical deployment is toward smaller systems rather than larger ones.

 

Cost
The obvious force. Running a large model is expensive per request, and at volume that cost dominates everything.

 

For a person asking a few dozen questions a day it does not matter. For a product handling millions of requests, or a workflow processing every document a company receives, it is the whole economics.

 

What changed is that smaller models became good enough for a wide range of tasks. Classification, extraction, routing, summarising, straightforward drafting: for these, a small model frequently produces work indistinguishable from a large one, at a fraction of the cost and considerably faster.

 

The pattern this produces is routing. Use a small model by default, escalate to a large one when the task genuinely needs it. Most requests do not.

 

Privacy
The second force, and for regulated industries the decisive one.

 

A model running on your own infrastructure, or on the device itself, means data does not leave. That resolves a category of problem that is otherwise negotiated contract by contract: data residency, retention terms, whether inputs are used for training, what happens in a vendor breach.

 

For healthcare, legal, financial services and public sector work, this is often not a preference. It is the condition of using the technology at all.

 

Small models are what make this viable. Running a very large model on your own hardware is an expensive infrastructure project. Running a small one is increasingly ordinary.

 

Latency
The least discussed and, for certain uses, the most limiting.

 

A round trip to a data centre takes time. Usually not much, but enough to matter for anything interactive: autocomplete as someone types, live transcription, responsiveness in an interface where a pause is noticeable.

 

Local execution removes the round trip entirely. It also removes the dependency on connectivity, which matters more than people in offices tend to assume. Field work, aviation, shipping, anywhere with unreliable network.

 

What this does not mean
Frontier models are not going away, and treating this as a replacement narrative gets it wrong.

 

The hardest work still needs the most capable systems: complex reasoning, long documents held in view at once, tasks where quality genuinely dominates cost. That will continue, and the frontier will keep moving.

 

The realistic picture is a split rather than a winner. Large models for the hard fraction, small models for the volume, with routing between them. Most well-built systems already work this way even where users never see it.

 

What to take from this
If you are building something, do not default to the largest available model. Test whether a smaller one handles your actual task. The saving is often substantial and the quality difference frequently undetectable for narrow work.

 

If a privacy constraint has been blocking you, check the position again. What was infeasible on your own infrastructure two years ago may be routine now, and the constraint may no longer be binding.

 

If you are buying, ask what the vendor runs underneath and whether it is fixed. A tool locked to one expensive model has a cost structure that will not improve. One that routes intelligently will get cheaper as smaller models get better.

 

The thing to watch
The interesting question is not how small a model can get. It is how narrow a task can be and still be worth a dedicated model.

 

The trend is toward specialised smaller systems trained for a particular domain, outperforming much larger general ones within that domain. If that continues, the future looks less like one enormous system answering everything and more like many modest ones, each good at something specific, with something coordinating between them.

 

That is a considerably less dramatic picture than the one usually painted, and it is closer to how most successful technology has actually developed.

AI forecasting is a saturated market with almost no accountability. Nobody revisits last year's predictions, so there is no cost to being confidently wrong.

 

Predictions Worth Ignoring, and the Few Worth Watching

 

A filter helps. Four things that reliably indicate a prediction is not worth your attention.

 

It has a date and no mechanism. "By 2028, AI will handle sixty percent of customer service." Where does the sixty come from? What has to happen first? A prediction with a precise number and no causal chain is a vibe wearing a suit.

 

It extrapolates a straight line. Two years of rapid improvement projected forward indefinitely. Technologies have S-curves, plateaus and bottlenecks. Straight-line extrapolation is the single most common error in this space.

 

It treats a job as a task list. "This will replace accountants." Accounting is a task list plus judgment, client relationships, professional liability and regulatory accountability. Predictions about roles that only address the tasks are addressing a fraction of the thing.

 

The predictor benefits from it being believed. Not disqualifying on its own. Worth weighting, particularly when the prediction is raising money.

 

What is actually worth watching
Three signals, all boring, all more informative than any forecast.

 

Cost per unit of capability. What it costs to do a given task has been falling steadily. That trend line determines what becomes economically viable, and it is measurable rather than speculative.

 

Reliability on multi-step tasks. Single-response quality is largely solved for most business purposes. Whether systems can chain ten steps without compounding errors is the open question, and it gates most of the interesting applications.

 

What regulated industries actually deploy. Healthcare, financial services and legal move slowly because the consequences of error are real. When they deploy something in production, that is evidence of dependability in a way that no benchmark provides.

 

The useful stance
Assume the direction is right and the timeline is wrong.

 

Most confident predictions about AI have the direction roughly correct and the speed badly off, usually too fast in the short run and too slow over longer periods. That is the historical pattern for most significant technologies.

 

Which suggests a practical posture: build for where things are now, stay aware of the direction, and avoid making commitments that depend on a specific capability arriving by a specific date.

 

The people who got the last technology wave right were rarely the best forecasters. They were the ones who noticed what was already working and acted on it.

Amazon Bets on Rat Brains to Make AI Video 5x Faster

 

A Startup Lands on Amazon's Shelf
On September 22, 2026, The Biological Computing Co. (TBC), an applied biological computing company that uses real neurons to improve AI models, announced a collaboration with AWS to bring what it calls the world's first neuron-derived AI video model to market. Headlines have called these AWS "contracts," but the public record points to a partnership and distribution deal. Contract values and terms haven't been publicly confirmed. It still matters. TBC plans to run its model on AWS Trainium, make it deployable through Amazon SageMaker AI and pursue distribution in AWS Marketplace. For a young company in a once-fringe field, that's a direct line to a massive enterprise customer base. AWS and TBC expect access to reach all AWS enterprise customers, but haven't given a date.

 

What "Rat Brain-Powered" Actually Means
Biological computing uses living cells, usually neurons, to process or study information. TBC places living cells on multi-electrode silicon arrays made by 3Brain, a Swiss biotech company, working with rat brain neurons and human stem cells. Here's the twist most headlines skip. There are no rat neurons inside the product. The modified model runs on ordinary AI hardware, with no living cells on the customer's side. The biology stays in the lab and inspires the software.

 

The Workflow, Step by Step
TBC describes a loop. Researchers send electrical signals into the neurons, record how they respond and look in that activity for computational patterns that can be written as software. That software then sits on top of an existing video generator. TBC won't name the underlying open-source model beyond calling it a frontier model, and says its neuron-derived layer adds less than 0.1% to it. Exactly how neural recordings become code hasn't been made public.

 

Why Video Came First
The founders give two reasons. The electrode arrays are grids, and images could be mapped onto them more readily than text. And investor Jeff Dean reportedly pushed TBC to start with video because established benchmarks exist there. The payoff could be big. TBC claims 80% lower inference costs and 5x faster generation, with better quality than the base model. Cheaper, faster clips mean creators can iterate more and platforms can serve more users. The catch? The numbers are self-reported, and with the baseline kept secret, outsiders can't reproduce them yet.

 

The Stress Test Nobody Has Run
AWS itself raised the hard question. Deap Ubhi, who heads technology for startups at AWS, wants to see whether the gains hold on videos ten minutes or an hour long, where a model has to stay consistent with everything it already generated. That's AI video's sore spot. Objects drift, faces change, scenes forget themselves. If neuron-derived tweaks help there, it's a real breakthrough. If they only shine on short clips, the story gets smaller.

 

What It Means for the Future of AI
As a tech fan, I find this genuinely exciting. The human brain runs on roughly 20 watts, while AI compute bills keep climbing, so mining biology for efficiency ideas is a smart hedge. TBC also has company. AWS works with Cortical Labs, an Australian company that combines lab-grown neurons with silicon chips. The risks deserve equal airtime. Human stem-cell neurons raise questions about consent, sourcing and how cultures are handled. No jurisdiction currently regulates commercial biological computing as a category. And nobody has shown yet that the biology is essential, or whether conventional optimization could find the same shortcuts.

 

Verdict: Weird, Promising, Unproven
TBC's pitch is refreshingly practical. Ubhi says the startup is working within existing generative AI standards and asking how to make current visual models more efficient. That's what caught Amazon's eye. Now it needs independent benchmarks. If the numbers hold, lab-grown rat neurons may quietly shape how the next wave of AI video gets made.

 

Sources

 

Amazon Press Center, "The Biological Computing Co. Partners with AWS to Bring World's First Neuron-Derived AI Video Model to Market":

 

https://press.aboutamazon.com/aws/2026/9/the-biological-computing-co-partners-with-aws-to-bring-worlds-first-neuron-derived-ai-video-model-to-market
The Biological Computing Co. official site:

 

https://www.tbc.co/

 

Related YouTube Videos
"In the Field: The exclusive tour inside TBC.co's lab | Amazon Web Services":

 

https://www.youtube.com/watch?v=R4eKRihw9f4
"This computer runs on living human brain cells | REUTERS":

 

https://www.youtube.com/watch?v=3KeC8gxopio