I spent this month doing what I always do - reading, tracking, questioning. And somewhere between the headlines screaming about billion-dollar chip deals and the quieter stories about India mandating AI labels, I realized we’re living through something strange. We’re witnessing both the hyper-acceleration of AI’s technical capabilities and the sudden, urgent attempt by governments worldwide to grab the reins. The technology is sprinting and as always policy is trying to catch up.
Let me walk you through what happened this month. Not just the what, but the why it matters, the parts most people missed, and the questions we should be asking as we barrel toward 2026.
The Browser Wars: OpenAI’s Atlas and the Fight for Your Attention
On October 21, OpenAI launched ChatGPT Atlas - a web browser with AI baked into its core. It’s available on macOS now, with Windows, iOS, and Android versions coming soon. The pitch? A “true super-assistant” that understands your browsing context, remembers what you’ve explored, and can complete tasks for you without you having to copy-paste or leave the page.
It sounds convenient. It probably is convenient. But here’s the thing nobody’s saying out loud: this isn’t about making your life easier. It’s about data.
OpenAI needs scale. ChatGPT is impressive, but to keep improving, these models need to see how humans actually navigate the web - not just what they type into a chat box. Atlas gives OpenAI a window into every click, every search, every hesitation you have while booking a train ticket or comparing hotel prices. The “browser memories” feature? Optional, yes. But it’s storing context from sites you visit for 30 days on OpenAI’s servers.
And if you’re a paying subscriber ($20/month for Plus, more for Pro), you get “agent mode” - the AI can actually do things for you, like book appointments or research topics while you browse. Free users hit message limits quickly. The game here looks to be, get you hooked on convenience, then monetize through subscriptions or, eventually, something else.
But actually, Atlas might fail and yet it would still win on something. Google Chrome has 71% market share. People don’t switch browsers easily - there’s inertia, there are saved passwords, there’s muscle memory. OpenAI is betting you’ll abandon all that for a chatbot in your sidebar. But here’s something worth pondering, even if Atlas flops as a product, it succeeds as a data collection engine. Every user who tries it is teaching OpenAI how humans think through web tasks. That data feeds back into GPT models. It’s not about winning the browser wars. It’s about not being left behind in the intelligence race.
Microsoft responded two days later by expanding Copilot Mode in Edge with similar AI actions and journeys. The fight isn’t for browser dominance. It’s for behavioral data at scale.
The Chip Wars Heat Up: AMD Breaks Nvidia’s Stranglehold
If October had a single deal that reverberated across the entire AI infrastructure landscape, it was this: AMD and OpenAI announced a multi-year partnership to deploy 6 gigawatts of AMD GPUs, starting with 1 gigawatt in the second half of 2026.
Six. Gigawatts. To put that in perspective, that’s enough computing power to run cities.
But the deal had a kicker: OpenAI received warrants to potentially buy up to 10% of AMD - about 160 million shares—at just one cent per share, contingent on hitting deployment milestones. AMD’s stock surged 34%, adding $80-100 billion in market value in a single day. This partnership is expected to generate tens of billions of dollars in revenue for AMD over the next five years.
Why this matters: Nvidia has had a near-monopoly on AI chips, controlling over 90% of the market. That dominance has created bottlenecks - companies like OpenAI, Google, and Meta have been at the mercy of Nvidia’s supply constraints and pricing. AMD’s deal breaks that stranglehold. It signals that OpenAI is diversifying its compute supply chain, which is smart. Relying on a single vendor in a sector moving this fast is a risk no one wants to take.
But… this isn’t really about chips. It’s about power, literal electrical power. Training and running these massive AI models requires ungodly amounts of energy. Six gigawatts is the equivalent of several nuclear power plants. As AI data centers proliferate, we’re running into a new constraint: not silicon, but electricity and cooling capacity. Meta announced a $1.5 billion data center in Texas this month for AI workloads. The conversation is shifting from “do we have enough chips?” to “do we have enough power?”
And then there’s Qualcomm, jumping into the fray on October 27 with its AI200 and AI250 chips aimed at data centers. These chips are designed for AI inference (running models, not training them) and promise significantly lower power consumption. Qualcomm’s entry - along with Intel’s new Crescent Island chip, means we’re entering a period of genuine competition. That should, theoretically, push performance up and costs down. But it also fragments the ecosystem. Different chips, different architectures, different optimization requirements. The standardization battles are going to get messy.
But here is what many people missed: Ray Dalio, the legendary investor, said on October 28 that an AI market bubble is forming, but it probably won’t pop until the Federal Reserve tightens monetary policy. His “bubble indicator” is high. He’s not wrong. AMD’s market cap jumped by tens of billions on a deal that won’t deliver revenue for years. Tesla, Nvidia, tech giants valuations are frothy. But bubbles don’t pop on fundamentals alone. They pop when liquidity dries up. As long as the Fed keeps rates accommodative, the AI hype machine can keep running. The question isn’t if there’s a bubble. It’s what will make it pop?
The Regulation Wave: From Brussels to Beijing to Sacramento
If September was about building, October was about governing. And boy, did governments show up.
Europe: The AI Act Enters Implementation Phase
On October 8, the European Commission published its “Apply AI Strategy” - a comprehensive policy framework to accelerate AI adoption across 11 strategic sectors, including healthcare, robotics, manufacturing, defense, energy, and public services. This isn’t just regulation. It’s industrial policy. Europe is trying to position itself as the global leader in trustworthy AI, not just any AI.
The strategy includes sectoral “flagships” with specific actions and target dates laid out through 2026 and beyond. The EU AI Act itself - the world’s first comprehensive AI legal framework continues its phased rollout. As of October 2025, high-risk AI systems must complete risk assessments. By December, they must register in the EU database. And by mid-2026, full compliance kicks in.
On September 26, the Commission opened public consultation on draft guidance for reporting serious AI incidents under Article 73 of the AI Act - basically, when your AI system screws up in a way that could harm people, you have to tell authorities. The guidance includes practical examples: misclassifications, accuracy drops, system downtime. It’s technical, it’s detailed, and it shows Europe is serious about enforcement.
Italy became the first EU member state to adopt national legislation complementing the AI Act—Law No. 132, which entered force on October 10. It establishes sector-specific rules for healthcare, public administration, judicial activity, national security, and employment. The framework maintains definitional consistency with the EU Act while addressing Italy’s specific needs, including dual-tier consent for minors (under 14 requires parental consent, 14-18 can consent themselves).
The edge angle: Europe is building the world’s most sophisticated AI regulatory architecture. But regulation isn’t the same as innovation. The U.S. and China are racing ahead on deployment and application. Europe risks becoming the world’s AI regulator while Silicon Valley and Shenzhen build the actual AI economy. The Apply AI Strategy tries to fix this by pushing adoption, but cultural and structural barriers remain. European venture capital is risk-averse. Labor protections make rapid iteration harder. The question is whether Europe can have both: strong governance and competitive industry. History suggests that’s a tough balance.
United States: California Goes Big, Federal Government Goes… Lighter
California Governor Gavin Newsom signed the Transparency in Frontier Artificial Intelligence Act on September 29, making it the nation’s first comprehensive AI safety law. It applies to developers of “frontier models” - foundation models trained with more than 10^26 operations (basically, the biggest, most powerful models).
Requirements include:
Publishing safety frameworks on company websites
Reporting “critical safety incidents” that result in physical harm
Whistleblower protections for employees flagging catastrophic risks
But here’s where it gets interesting, the final version was heavily watered down from the original bill. First-time violation penalties dropped from $10 million to $1 million. Incident reporting only covers physical harm - not financial damage, privacy breaches, or other non-physical harms. For billion-dollar companies, a $1 million fine is a rounding error. Critics argue it creates a “cost-benefit calculation” where safety violations become just another line item. ( Leave your take below )
Still, California passed 18 new AI-related laws in 2025. The direction is clear: states are filling the federal vacuum. Meanwhile, the Trump administration’s “America’s AI Action Plan”, released in July, focuses on removing regulatory barriers and accelerating innovation. The Executive Order 14179 from January 2025 explicitly revoked Biden’s 2023 Executive Order on AI safety, calling it impediments to U.S. dominance.
What people are missing: Federal-state divergence is creating a patchwork regulatory environment. California’s rules are de facto national rules because companies aren’t going to build separate products for different states. But when federal policy actively opposes state regulation, you get legal uncertainty. Companies don’t know what the rules will be in 18 months. That uncertainty itself becomes a brake on innovation, which is ironic given the federal government’s stated goal.
China: The AI+ Campaign and Global Governance Ambitions
China released its Global AI Governance Action Plan on July 26 at the World AI Conference 2025. It proposes a 13-point roadmap covering innovation, infrastructure, open ecosystems, data sharing, standards development, and international cooperation. China also proposed establishing a global AI cooperation organization, potentially headquartered in Shanghai, to foster international collaboration and prevent monopolistic control by a few countries or corporations.
In August, China issued the “AI Plus” initiative implementation guideline, targeting a 70% penetration rate of intelligent terminals and AI agents by 2027, and 90% by 2030. In October, the Ministry of Industry and Information Technology released draft Administrative Measures for the Ethical Management of AI Technology, requiring organizations to establish AI ethics committees and conduct ethics reviews for AI projects that pose ethical risks.
China also rolled out AI labeling rules. Starting in October, AI-generated content providers must display clear labels to identify material created by artificial intelligence. It’s similar to what India proposed (more on that below), but China moved faster.
Chinese AI models are dominating global rankings. Chinese models occupy 9 of the top 10 positions on Hugging Face, the leading open-source AI community. DeepSeek’s R1 model, which cost only $294,000 to train (on top of the $6 million base model), rivals OpenAI’s o1 in reasoning tasks. Daily token consumption in China went from 100 billion at the start of 2024 to over 30 trillion by mid-2025 - a 300-fold increase.
The contradiction: China is simultaneously pushing aggressive AI deployment domestically while advocating for global governance frameworks internationally. It’s a smart play. Position yourself as the responsible actor calling for multilateral cooperation while the domestic industry races ahead under state support. The U.S. has historically done something similar (call for open markets while protecting strategic industries). The difference is China’s state-driven industrial policy is more explicit, more coordinated, and, arguably, more effective at achieving specific technology outcomes in compressed timeframes.
India: Labeling Deepfakes in a Sea of Misinformation
On October 22, India’s Ministry of Electronics and Information Technology proposed amendments to the IT Rules, 2021, mandating clear labeling of all AI-generated content across social media platforms.
The proposed rules are strict:
For visual content, labels must cover at least 10% of total display area
For audio content, labels must be audible during at least 10% of total duration
Permanent metadata identifiers or watermarks must be embedded
Platforms must obtain user declarations at upload time regarding whether content is AI-generated
Platforms must deploy automated detection tools to verify declarations
Both creators and platforms are responsible. Failure to comply could result in platforms losing safe harbor immunity.
India has nearly 1 billion internet users. In a diverse nation with multiple ethnic and religious groups, misinformation - especially during elections, can incite violence. Authorities have raised concerns about AI-generated deepfake videos featuring political figures. The rules aim to “guarantee clear labeling, metadata traceability, and transparency for all publicly accessible AI-generated media”.
The catch: Enforcement. How do you verify compliance across millions of daily uploads on Instagram, YouTube, WhatsApp, and X? Automated detection tools aren’t perfect, AI-generated content is getting harder to distinguish from human-created content. And what about satire, parody, or artistic expression? Where’s the line?
The creative industry is already pushing back, calling the 10% label rule “overreach”. They argue it destroys the aesthetic value of digital art and content creation. There’s a fundamental tension between transparency and creative freedom.
What’s being overlooked: This is a preview of global regulation to come. The EU has similar requirements under the AI Act. China has labeling rules. If the U.S., Europe, India, and China all move toward mandatory AI content labeling, it becomes a de facto global standard - not because of international agreement, but because platforms can’t manage dozens of different regional systems. They’ll default to the strictest common denominator.
But here’s the darker possibility: what if labeling normalizes synthetic content? Right now, a deepfake feels deceptive. But if every AI-generated image and video is labeled, and half the content online is labeled, do we just… get used to it? Does the label become wallpaper? I’m not sure we’ve thought through the second-order effects.
The Product Blitz: Sora 2, Veo 3.1, and the Video Generation Explosion
October was also the month video generation went mainstream.
OpenAI officially launched Sora 2 on October 1, marking a quantum leap in text-to-video generation. The new model generates 60-second videos with cinema-quality resolution, improved physics understanding, enhanced temporal consistency, and for the first time, integrated high-fidelity, context-aware audio generation. You don’t just get the visuals. You get synchronized sound effects, dialogue, ambient noise.
The Sora iOS app hit 1 million downloads in under five days, faster than ChatGPT’s initial debut. OpenAI also introduced a “cameo” feature, allowing users to insert their own likeness and voice into generative videos.
But it immediately sparked controversy. Hollywood studios raised copyright backlash over the use of protected characters and voices in training data. It’s the same battle playing out in every creative industry: Who owns the training data? Who gets compensated? What’s fair use in the age of generative AI?
Google responded on October 15 by releasing Veo 3.1 with native audio support, improved prompt adherence, and granular editing controls Veo 3.1 is available in Google’s Flow video editor, the Gemini app, and via the Gemini API in Google AI Studio and Vertex AI. You can now guide video generation with up to three reference images for character consistency across shots, extend existing clips, or generate transitions between a first and last frame.
Google says that since Flow’s launch in May, users have created more than 275 million videos. That’s staggering. We’re not talking about hobbyists anymore. Video generation at scale has arrived.
The implication: We’re entering the era of synthetic media abundance. Creating a 60-second video used to require a production team, expensive equipment, weeks of editing. Now it takes a text prompt and 30 seconds. What happens when anyone can create Hollywood-quality video content? The optimistic view: democratization of creativity, explosion of new storytelling forms. The pessimistic view: deepfake chaos, misinformation at scale, erosion of trust in any visual media.
I think the truth is both. And I think we’re not ready.
The part people are ignoring: Energy consumption. Training Sora 2, Veo 3.1, and similar models requires massive computational resources. Running inference at scale - especially 60-second videos with audio for millions of users - requires even more. Data centers are already consuming at least 3% of global power, projected to hit higher percentages by 2030. Video generation accelerates that curve. We’re trading carbon for pixels.
At some point, the environmental cost of synthetic media becomes a policy question, not just a technical one.
The Safety Debate: 850 People Call for a Superintelligence Ban
On October 22, the Future of Life Institute released a 30-word statement calling for a prohibition on the development of superintelligence until there is “broad scientific consensus that it will be done safely and controllably, and strong public buy-in”.
Over 850 public figures signed it, including:
AI pioneers Geoffrey Hinton and Yoshua Bengio (the “Godfathers of AI”)
Apple co-founder Steve Wozniak
Virgin Group founder Richard Branson
Former royals Prince Harry and Meghan Markle
Conservative commentators Steve Bannon and Glenn Beck
Nobel laureates, evangelical leaders, and policymakers
The coalition is bizarre and fascinating - tech visionaries, liberals, conservatives, royals, priests. What unites them? Fear. Fear of what superintelligence - AI that significantly outperforms all humans on essentially all cognitive tasks, might do if developed recklessly.
The statement warns of concerns ranging from “human economic obsolescence and disempowerment, losses of freedom, civil liberties, dignity, and control, to national security risks and even potential human extinction”.
This isn’t the first such letter. In March 2023, FLI organized a letter calling for a six-month pause on training powerful AI systems. That didn’t happen. So why try again?
Anthony Aguirre, FLI’s executive director, told TIME that they believe superintelligence could arrive in as little as one to two years. “Time is running out,” he said. The only thing likely to stop AI companies from barreling toward superintelligence is “widespread realization among society at all its levels that this is not actually what we want”.
The contrarian view: Calls for bans rarely work. Companies like OpenAI, Anthropic, Google, and Meta have billions invested. National governments see AI as a strategic technology race - China vs. U.S., with Europe trying to carve a third path. No one wants to be the country that blinks first and falls behind.
The statement also has a fundamental problem: it’s vague. What is “broad scientific consensus”? Who determines when we’ve achieved “strong public buy-in”? These are political questions disguised as technical ones. There’s no neutral arbiter. The UN? Too slow. The U.S. government? Too partisan. The EU? Lacks enforcement power beyond its borders.
But here’s what I keep thinking about: the letter’s signatories include Steve Bannon and Prince Harry. That’s not a coalition you see every day. When people with radically different worldviews agree that something is a threat, maybe, just maybe - it’s worth slowing down and asking hard questions before we cross lines we can’t uncross.
The Hardware Subplot: Meta’s Layoffs, Qualcomm’s Bet, Tesla’s Robots
While the world obsessed over models and policies, the hardware story kept churning beneath the surface.
Meta Cuts 600 AI Jobs
On October 22, Meta announced roughly 600 layoffs in its Superintelligence Labs AI division - about 10-15% of the AI workforce. Meta described the cuts as a step to make the AI unit “more flexible and responsive,” but it came just months after a massive AI hiring spree that cost hundreds of millions and included bringing in Scale AI’s CEO, Alexandr Wang, as Meta’s new chief AI officer.
Why the cuts? Meta is trying to streamline. The company reorganized its AI teams under Meta Superintelligence Labs in June, but the Llama 4 model received a lukewarm reception. Meanwhile, Meta committed $15 billion in a partnership with Scale AI. The layoffs suggest internal course correction, shedding bureaucracy, focusing resources.
Meta is also integrating AI deeply into its ad platform. Starting December 16, Meta will use what people type or say to Meta AI - including text and voice to personalize content and ads on Facebook, Instagram, and other platforms. Over 1 billion people use Meta AI monthly. Your conversations with the chatbot become signals in the recommendation engine. If you chat about hiking, you’ll see hiking groups, hiking posts, ads for hiking boots.
You can’t opt out. The only way to avoid it is not to use Meta AI. And it won’t apply in the EU, UK, or South Korea at launch due to regulatory considerations.
Translation: where privacy laws are strong, Meta holds back. Everywhere else, it’s open season on your conversational data.
Tesla’s Optimus: From Lab to Times Square
Tesla’s Optimus humanoid robot made multiple public appearances in October. On October 27, Optimus handed out candy in Times Square. Tesla Board Chair Robyn Denholm revealed that Optimus can now fold laundry, wipe tables, and shake hands. “The tactile nature of his hand is actually really very good,” she told CNBC.
During Tesla’s Q3 earnings call on October 26, Elon Musk said that Optimus V3 will be unveiled in Q1 2026. He described it as looking so lifelike “you’ll need to poke it to believe that it’s actually a robot”. Musk also said Optimus has the potential to be “the biggest product of all time” and could perform delicate tasks like surgery in the future.
Musk’s vision: “With Optimus and self-driving, we can actually create a world where there is no poverty, where everyone has access to the finest medical care”.
Reality check: Optimus is impressive, but it’s not close to mass production. The robots at Tesla’s “We, Robot” event in October 2024 relied heavily on teleoperation - humans controlling them remotely. The company hasn’t been transparent about how much is autonomous vs. tele-operated.
Musk also announced in March 2025 that an Optimus robot would be sent to Mars in 2026 aboard a SpaceX Starship. That timeline is… optimistic. But the point isn’t whether Musk hits his timelines. The point is Tesla is the only major automotive company investing heavily in humanoid robotics and general-purpose AI at the same time. Ford, GM - they’re nowhere on this.
If Optimus works, it changes everything: manufacturing, elder care, hospitality, logistics. If it doesn’t, Tesla spent billions building advanced toys. Either way, we’re watching one of the most audacious technology bets of the decade play out in real time.
Microsoft and Anthropic: Memory Wars
Microsoft unveiled 12 major Copilot updates in its Fall 2025 release on October 23. The highlights:
Groups: Collaborative Copilot sessions for up to 32 participants
Imagine: A creative hub for generating and remixing AI content
Mico: A new character interface (think Clippy for 2025)
Real Talk: A conversational mode with constructive pushback
Memory & Personalization: Long-term memory of user preferences, dates, goals
Copilot Mode in Edge: AI-driven summarization, comparison, and web actions
Anthropic rolled out its automatic “memory” feature to all Claude Pro and Max subscribers on October 23. Claude can now recall past conversations without being explicitly asked. Users can view, edit, or delete what Claude remembers, and each “Project” in Claude has separate memory spaces to keep work and personal chats distinct.
ChatGPT, Gemini, and now Claude all have memory. It’s table stakes. The competition is converging on the same feature set, which means differentiation will come from execution quality and ecosystem lock-in, not novel capabilities.
The unspoken trade-off: Memory makes AI more useful. But it also makes it more invasive. These systems are learning your habits, preferences, anxieties. That data is valuable to you, to the companies, and potentially to third parties. Where is it stored? How long? Who has access? The answers vary by provider and jurisdiction. Most users don’t read the terms of service. They just click “yes” and hope for the best.
What All This Means: Five Trends to Watch
As I sit here at the end of October, trying to make sense of the chaos, a few patterns emerge.
1. The Geopolitical AI Race Is Heating Up, Not Cooling Down
The U.S., China, and Europe are pursuing fundamentally different strategies. The U.S. is betting on private-sector innovation with light-touch regulation. China is combining state-directed industrial policy with rapid deployment and global governance positioning. Europe is building the world’s most sophisticated regulatory framework while trying (and struggling) to keep pace on innovation.
None of these approaches is objectively “right.” They reflect different values, different risk tolerances, different political economies. But the divergence means there won’t be a single global AI governance regime. Instead, we’ll get regional blocs with different rules, different norms, different red lines. Companies operating globally will have to navigate that fragmentation. So will policymakers.
2. Infrastructure, Not Models - Is the New Bottleneck
The AMD-OpenAI deal, the Qualcomm chip launch, Meta’s $1.5 billion Texas data center - these aren’t just hardware stories. They’re about recognizing that scaling AI requires physical infrastructure: chips, data centers, cooling systems, and, most critically, electrical power.
Six gigawatts for a single partnership. That’s more power than some countries use. We’re hitting the limits of existing grid capacity in some regions. The AI buildout will force energy infrastructure upgrades, which will take years and billions in investment. Or we’ll hit a wall where the models stop scaling because we literally can’t power them unless we have more optimization breakthroughs.
This also has climate implications. AI is energy-intensive. If we don’t green the grid, AI expansion accelerates carbon emissions. That’s not a hypothetical - it’s math.
3. Regulation Is Coming, But It’s Patchwork and Reactive
California’s SB 53, India’s AI labeling rules, Europe’s AI Act, China’s ethics measures - governments are moving. But they’re moving in different directions, at different speeds, with different enforcement mechanisms.
The result is regulatory fragmentation. Companies face compliance burdens across jurisdictions. Some places have strong rules (EU), some have weak rules (U.S. federal), some have contradictory rules (federal vs. state in the U.S.). That creates uncertainty, which ironically - might slow innovation more than the regulations themselves.
The other problem: regulation is reactive. Policymakers are trying to govern technologies they don’t fully understand, in industries moving faster than legislative cycles.
By the time a law passes, the technology has evolved. It’s a treadmill with no off switch.
4. Trust Is the Emerging Crisis
Deepfakes, synthetic media, AI-generated misinformation - October gave us a glimpse of what happens when the tools to create deceptive content become cheap, fast, and accessible. India’s labeling rules are a response to deepfake-driven election interference. Europe’s AI Act incident reporting requirements are about accountability when systems fail.
But labels and laws can’t fully solve the trust problem. If every image, video, and audio clip might be synthetic, how do we know what’s real? We’re entering an era where “seeing is believing” no longer holds. That has profound implications for journalism, justice, democracy, and social cohesion.
Some argue cryptographic provenance - embedding verification into media at the point of creation could help. Maybe. But that requires universal adoption, which requires coordination, which requires… well, we’re back to governance.
5. The Safety vs. Speed Debate Isn’t Resolved - It’s Intensifying
The superintelligence ban letter, California’s safety law, Ray Dalio’s bubble warnings, Meta’s layoffs - all point to growing tension between moving fast and moving carefully.
Industry wants to ship. Safety advocates want to slow down. Investors want returns. Policymakers want votes. There’s no neutral arbiter. No one has the authority to call time-out.
What worries me most isn’t that we’re moving fast. It’s that we’re moving fast while pretending we’ve thought through the consequences. We haven’t. We’re making it up as we go, hoping the upsides outweigh the downsides, trusting that if things go wrong, we’ll figure it out.
Maybe we will. History has a way of muddling through. But October 2025 was a month when the warning lights started flashing brighter. Whether we listen is another question.
Closing Thoughts: October as Prologue
Here’s what I keep coming back to: October felt like a month of contradictions.
We saw unprecedented technical progress—Sora 2 generating 60-second videos with audio, AMD securing a 6-gigawatt chip deal, Chinese models dominating global rankings. And we saw governments scrambling to impose rules before losing control entirely.
We saw billion-dollar valuations and bubble warnings. We saw open-source breakthroughs . We saw calls for superintelligence bans signed by 850 people, and we saw companies hiring video game tutors to train chatbots.
It’s thrilling and terrifying in equal measure.
As someone who spends their days thinking about AI policy - reading the white papers, tracking the regulatory proposals, watching the geopolitical chess moves, I’m struck by how much we don’t know. We don’t know if the AI bubble will pop next year or in 2030. We don’t know if Europe’s regulatory approach will become the global standard or a cautionary tale. We don’t know if superintelligence will arrive in two years or twenty, or if it will look anything like we expect.
What we do know is this: October 2025 was a pivot point. The month when AI stopped being a niche technology concern and became a mainstream governance challenge. The month when chips became as strategic as oil. The month when “synthetic media” moved from research papers to election misinformation to regulatory mandates.
November and December will bring more. CES in January 2026 will bring more. The question isn’t whether AI will keep accelerating. It will. The question is whether our institutions - governments, companies, civil society, academia, can build the guardrails, norms, and accountability mechanisms we need before the technology runs too far ahead.
I don’t have the answer. But I do know we need to be asking the question. Loudly, persistently, inconveniently.
Because if October taught us anything, it’s that the future doesn’t wait for permission. It just arrives, ready or not.
Samuel Abinsinguza writes on AI governance and policy. He’s currently building CentPol (Center for Emerging and Next Tech) and thinking about how developing economies can participate in, and shape the AI future. Follow for more musings on technology, policy, and the spaces in between.
The Policy Brief is a timely newsletter making sense of AI’s intersection with governance, society, and power. Subscribe for the Next edition.