Headlines this week - Apr 6, 2025
A look at how capital is being deployed across future opportunities
This week in the future:
AI safety is turning into a priority, as models become more powerful and more widely adopted
We have confirmation that the current AI models pass the Turing Test. In a paper published this week, researchers have found that participants chatting with a human and, simultaneously, with OpenAI’s GPT-4.5, believed the model was a human 73% of the time. Actually, in an X thread, one of the researchers claims that “GPT-4.5 was even judged to be human significantly more often than actual humans”
Google DeepMind wants us to anticipate the era of Artificial General Intelligence (AGI), and the associated risks. This week they published a long paper describing their approach to “Technical AGI Safety and Security”, in which they claim to be “prioritizing readiness, proactive risk assessment, and collaboration with the wider AI community”
The company is optimistic, but calls for responsibility, too. They believe this technology can be a catalyst for progress, but also that its power requires the assessment and prevention of any potential harms that it might cause. The paper explores four main areas: (1) Misuse, (2) Misalignment, (3) Mistakes, and (4) Structural risks (linked to multi-agent dynamics, where no single agent is at fault)
In parallel, Helen Toner, an ex-board member at OpenAI, has published a white paper on this very same topic. She expresses her concern that “human-level AI” could possibly be delivered in the 2030s, and claims that, “we as a society are not ready to handle human-level AI and all its implications that soon”. Her point is that, even if these are only possibilities, not guaranteed outcomes, we should start preparing (and she makes specific proposals for that in her article)
Some futurists are seeing a rather gloomy scenario for the coming years. The “AI Futures Project”, a nonprofit led by an ex-OpenAI researcher, based in Berkeley, has released a report and a website called “AI 2027” presenting a detailed fictional scenario of what could happen with AI in the next 2-3 years. And it does not look so good. Among other things, the picture assumes there will be fully autonomous AI agents, better than humans “at everything”, by the end of 2027. On top of that, the scenario shows these agents completely disrupting the global order…
On a related topic, concerns are growing on the future role / purpose of humans, in an “AGI world”
There is a positive perspective, if we realize that intelligence is not the only factor that makes us human. In a post this week, my ex-colleague X Amatriain (now at Google) claims that empathy, creativity and the ability to connect and feel are at least as important as intelligence, when looking at what characterizes a human being. So rather than trying to compete with AI systems in “raw intelligence”, he proposes to leverage their capabilities to enhance our own intelligence and promote the other, uniquely human aspects
But for now, we don’t seem to be paying attention to the “human factors”. An article at the FT looks at the “Studio Ghibli”-mania, last week, with millions of AI-generated memes turning people’s pictures into Japanese anime snapshots. Behind all this, the author detects a careless “belief” on AI as a technology with “magical powers”, and a lack of attention for the details in the real, genuine Ghibli material / art. So we would actually be skipping the actual “human factors” that X Amatriain talks about. And this could have negative consequences for society, as we expand the scope of AI tools
An additional pessimistic view is that “delegating” mental effort to AI systems could make us more “stupid”. An article at the WSJ discusses how AI chatbots are now offering an “easy path” for tasks like writing documents, and for some people this is becoming a permanent “temptation”, leading to lazier and weaker minds. As AI becomes more powerful, the trend is that people would lose creativity and become less intelligent
At the same time, we might be on the way to human-machine convergence, with fast progress being made in brain-computer interfaces
25 clinical trials are already under way. Two companies in the US (New York’s based Synchron and Elon Musk’s Neuralink) and one in China (Neuracle Neuroscience) are actively recruiting volunteers to get a brain implant that will connect their brain signals with a computer
The first use cases will be therapeutic. The plan is to use the devices in patients of paralysis, and enable them to control a computer with their thoughts
71 patients have already controlled a computer directly with their neurons, and play simple video games (Pong), move a robot arm, and (in a few cases) speak
Restoring speech for paralyzed people looks like a killer app. In the trials, producing voice messages has been possible for just a few patients. But this is an area of active research. This week we learnt about a research team at UC Berkeley that are using AI to solve the “latency problem” and make it possible for translate brain signals into audible speech in real time
As discussed in previous weeks, new “non-invasive” approaches could, if successful, accelerate progress. See the discussion on Alljoined, a startup using AI to simplify brain-computer interfaces,
Could renewable energies play a role in powering data centers?
Solar energy is a candidate, but its upsides should offset the downsides from the “intermittency problem”. The problem with solar, and other renewables, is the mismatch between the “continuity” in data centers’ energy demand and the “intermittency” on how these energies are produced. This leads to costly requirements of storage capacity (e.g. large batteries) that should be compensated by differential benefits that these energies bring
The number of solar deals in the pipeline seems to confirm that this is the case. According to a TechCrunch article this week, data center companies are moving to partner with solar energy companies. Since the beginning of 2025, 12 deals have been signed, each adding more than 100MW to the grid
The reports of a slowdown in demand for AI infrastructure might have been greatly exaggerated
The “Studio Ghibli” storm created an infrastructure bottleneck for OpenAI. This week Sam Altman went so far as to use social networks to call for help. On Tuesday at X, he admitted the company was having capacity problems and claimed to be looking for someone able to provide “GPU capacity in 100k chunks”
Coreweave’s stock rallied this week and is now above the IPO price. Coreweave, the AI infrastructure company that had a poor IPO performance on Friday last week, has now fully recovered from that. Even in the catastrophic market conditions after Trump’s tariffs, the stock has managed to close this week beyond the $40 IPO price, after rising +50% in just two days (before the storm). The takeaway is that there are still investors who believe in the growth of this business
Momentum about AI agents keeps growing, but there are still lots of uncertainties
AI agents could lead to “hybrid” workforces. Salesforce’s Marc Benioff is (as expected) fully bullish on AI agents, like the ones his company is building. He shared his vision in a WSJ article this week, in which he claims that (1) large models will be commoditized (he uses the DeepSeek to emphasize this), with value shifting to the apps (and the data), and (2) we’re moving into “hybrid” workforces, composed of humans and AI agents, and this will make it possible to increase productivity and free up human workers’ time for “higher-value” tasks (whatever this means…)
However, at the same time, there is still a lot of confusion about what to expect from an AI agent. An alternative view is that “agentic” has become a buzzword, that there is a hype around the concept, and that there is not yet a common view about what kind of “actions” would agents be expected to do on behalf of humans. Another WSJ article this week discussed these “contrarian” ideas
Amazon has seen an opportunity to simplify online commerce. Amazon is planning to launch “Nova Act” an AI agent that would aim to radically reduce friction for consumers in the e-commerce process. Act can carry out web searches, make purchases, or answer questions about what’s on the screen
Competition is heating up in the space industry
SpaceX is exploring new business models. This week the company launched a mission with four private astronauts, that orbited the Earth from pole to pole. This “space-tourism” service looks exactly like the kind of thing that Richard Branson used to dream about with his (mostly failed) company, Virgin Galactic. If successful, it could become a nice extra business for SpaceX
At the same time, competition is increasing:
In satellite communications, Amazon is emerging as a key rival. Amazon plans to launch the first batch of its “Kuiper” communication satellites (a direct rival to SpaceX’s Starlink) next week. 27 satellites are expected to be put in orbit on 9 April. Of course, there is a long way ahead, as Starlink has already 7,000 satellites in orbit, serving more 4.5m customers globally
In direct-to-device communications, Starlink is clashing with Apple. Both companies are competing for spectrum rights making it possible to connect rural areas, currently unattended by conventional cellular networks. Apple’s partner for satellite communication, Globalstar, is using SpaceX to launch the satellites, but Starlink needs Apple to connect its service to the iPhones
In rockets, the competitive race is also heating up. In this video by tech reporter Ashley Vance, C Kemp, the CEO of Astra Space, a startup building cheaper, more efficient rockets, describe the company’s plans to fly “Rocket-4”, their next-generation rocket
Researchers are working on the building blocks for space exploration. This week we learned about a team at University of Potsdam, Germany, that is developing a technique to build solar panels using lunar dust. This would open the door to self-sufficiency for a lunar base to build its own energy infrastructure
Investors are paying attention to AI’s promise to accelerate progress in biotechnology
Isomorphic Labs, a DeepMind biotech spin-off, has raised $600m. This is the first external funding round for this Google subsidiary, and it is among the biggest fundraisings for a UK AI company. Isomorphic Labs is behind the efforts, reported here in previous weeks, to take a new AI-generated drug to clinical trials by the end of this year. The round was led by Thrive Capital, also an OpenAI investor
A new wave of pharma M&A could increase the interest in biotech startups. Pharma companies are facing a “patent cliff”, with patents to “flagship” drugs expiring by the end of the decade, so they need new products in the pipeline, exactly what the new AI-driven biotech startups could provide. According to an FT article this week, only regulatory uncertainties in the US are stopping big pharma companies to start the new acquisitions wave. The perspective of startups being sold to large players could be part of what is exciting investors about the industry
China’s BYD consolidating as the global leader in electric vehicles, at the expense of Tesla
BYD’s unit sales increased +58% in 1Q2025 vs. 1Q2024. The company sold almost 1m new cars in the last quarter. Interestingly, growth was faster in hybrid models (+76%) than in pure electric ones (+39%) reflecting the point where demand is right now, including in China
Meanwhile, Tesla, under pressure in Europe, sold -13% less, and it does not look good. The figures missed analyst expectations and show how much the company is suffering in its reputation after Elon Musk’s political comments about some European countries (like Germany). Sales in Europe fell -76%. The perspectives are really bad, because sales in China (already falling -49%) could fall even faster with the trade war now being started by Trump
For leading AI companies, a tension is emerging between open sourcing foundational models and keeping things proprietary
OpenAI has decided to open source some of its models. The company, a clear representative of a “closed”, proprietary approach, has announced they will be releasing an “open-weight” language model. This would be similar to what Meta does with its Llama models, and would represent a shift in the policy that OpenAI has followed since at least 2019. Some analysts could also view it as a sign that “open source” has won (this was one of the key takeaways after the DeepSeek release some weeks ago)
At the same time, Google DeepMind will slowdown the pace of their research publications. Everyone understands this as a way to protect the company’s innovations from external parties copying them. This is pretty much an opposite move to what open source represents. But it also means that, even if the company open sourced some of its more “mature” models, they would still try to protect their “cutting-edge” ideas, and create a competitive advantage with them
LINKS:
1 - Population & natural resources
Biotech
Investors are paying attention to AI’s promise to accelerate progress in biotechnology
Isomorphic Labs, a DeepMind biotech spin-off, has raised $600m. Isomorphic Labs is behind the efforts, reported here in previous weeks, to take a new AI-generated drug to clinical trials by the end of this year. Google DeepMind’s drug discovery spin-off Isomorphic Labs raises $600mn
A new wave of pharma M&A could increase the interest in biotech startups. Pharma companies would need startups to address the “patent cliff” (patents to “flagship” drugs expiring by the end of the decade). And this should make those startups more attractive for investors. A biotech M&A revival is in the pipeline
Foodtech
A revolution in food production could be on the way, through the combination of AI and 3D printing. A Hong Kong research team aspires to enable “personalized food production” through this. HKUST Researchers Set to Transform Food Production with AI-enhanced 3D Food Printing Solution | The Hong Kong University of Science and Technology
Space
Competition is heating up in the space industry
SpaceX is exploring new business models. This could be the start of a “space tourism” business. SpaceX launches private astronaut crew in Fram2 polar-orbiting mission
At the same time, competition is increasing:
In satellite communications, Amazon is emerging as a key rival. Amazon plans to launch the first batch of its “Kuiper” communication satellites (a direct rival to SpaceX’s Starlink) next week. Amazon to begin deploying rival satellites to Starlink this month
In direct-to-device communications, Starlink is clashing with Apple. Apple’s partner for satellite communication, Globalstar, is using SpaceX to launch the satellites, but Starlink needs Apple to connect its service to the iPhones. Exclusive | Apple and Musk Clash Over Satellite Expansion Plans
In rockets, the competitive race is also heating up. Astra Space, a startup building cheaper, more efficient rockets, will fly its “Rocket-4”, their next-generation rocket (the CEO talks about it in this video). Chris Kemp on Astra's Comeback, Rockets and Drones from Space
Researchers are working on the building blocks for space exploration. E.g. this technique to build solar panels using lunar dust, that could make it possible for a lunar base to build its own energy infrastructure. Make solar panels out of the Moon to let us live in space, scientists say
Materials
A new wave of “meta-materials” is on the way, with good properties to build better robots. An European team is developing new materials which combine high stiffness with the ability to absorb and release large amounts of elastic energy. KIT - KIT - Media - Press Releases - PI 2025 - Metamaterials: Highly Twisted Rods Store Large Amounts of Energy
2 - Efficiency & Productivity
Energy
Nuclear
China is betting a lot of money on nuclear fusion. As an example, the Shanghai government is investing $1.4bn in China Fusion Corp., a local company looking to build a nuclear fusion reactor. China’s Megacity Shanghai Invests in Nation’s Fusion Energy Push
Renewables
Could solar energy play a role in powering data centers? It looks like the “intermittency problem” (the mismatch between the “continuity” in data centers’ energy demand and the “intermittency” on how these energies are produced) is not such a big obstacle. Data centers love solar: Here's a comprehensive guide to deals over 100 megawatts
New Transport Technologies
Electric Vehicles
China’s BYD consolidating as the global leader in electric vehicles, at the expense of Tesla
BYD’s unit sales increased +58% in 1Q2025 vs. 1Q2024. Growth was faster in hybrid models (+76%) than in pure electric ones (+39%) reflecting the point where demand is right now, including in China. BYD car sales soar as Tesla struggles in Europe
Meanwhile, Tesla, under pressure in Europe, sold -13% less, and it does not look good. Sales in Europe fell -76%. The perspectives are really bad, because sales in China (already falling -49%) could fall even faster with the trade war now being started by Trump. Tesla’s Global Vehicle Deliveries Sank 13% in First Quarter
Autonomous Cars
Public transport could be an attractive use case for autonomous vehicles, but only after removing some barriers. WeRide, a company building autonomous buses, claims that uncertainties in regulation and the ability to partner with local companies in each city are slowing down progress. China’s WeRide warns driverless tech profitability ‘difficult to predict’
Artificial Intelligence
AI: Apps
Agents
Momentum about AI agents keeps growing, but there are still lots of uncertainties
AI agents could lead to “hybrid” workforces. At least that’s what Salesforce’s Marc Benioff believes. Opinion | Say Hello to Your New Colleague, the AI Agent
However, at the same time, there is still a lot of confusion about what to expect from an AI agent. We don’t even know what kind of “actions” would agents be expected to do on behalf of humans. Everyone’s Talking About AI Agents. Barely Anyone Knows What They Are.
Amazon has seen an opportunity to simplify online commerce. Amazon is planning to launch “Nova Act” an AI agent that would aim to radically reduce friction for consumers with e-commerce. Amazon’s new AI agent is designed to do your shopping
Startups are working to make adoption easier. This is the case of Emergence AI, a company that has built a platform to facilitate AI agent creation at companies. Emergence AI’s new system automatically creates AI agents rapidly in realtime based on the work at hand
B2C
Counterintuitively, Apple and Amazon are perceived as AI laggards. Even if both companies were pioneers in “intelligent chatbots”, with Siri and Alexa, we seem to be waiting too long for a “GenAI” update to these apps. Apple and Amazon Promised Us Revolutionary AI. We’re Still Waiting.
B2B
The B2B market for GenAI is growing fast, according to Gartner. They expect GenAI spending to reach $644bn in 2025 (+76% vs. last year). Gartner forecasts gen AI spending to hit $644B in 2025: What it means for enterprise IT leaders
AI could work as a “natural selection filter” for companies. According to a recent study, there will be pain in the short-term, but companies that survive will be more sustainable long-term. If AI doesn't kill your company, it will make it stronger, study shows
Microsoft (like Google) looks bullish on AI’s role as a catalyst for scientific discovery. This comes after the recent launch by Google of specific models for this. Microsoft’s Christopher Bishop: Scientific discovery is AI’s killer application
Anthropic has launched a model specifically focused on education. This addresses some feedback they’ve collected on how students are currently using GenAI, and the observation that, as they currently are, models are not always so helpful. Anthropic flips the script on AI in education: Claude’s Learning Mode makes students do the thinking
AI: Robots
There is a battle under way between the US and China to dominate humanoid robots. Current efforts are focused on “embedding” large language models in physical bodies, a development that could represent a major leap forward in robots’ performance (and potentially in AI too). Humanoid Robots Are Lousy Co-Workers. China Wants to Be First to Change That.
AI: Foundational Models
For leading AI companies, a tension is emerging between open sourcing foundational models and keeping things proprietary:
OpenAI has decided to open source some of its models. The company, a clear representative of a “closed”, proprietary approach, has announced they will be releasing an “open-weight” language model. OpenAI to release open-source model as AI economics force strategic shift
At the same time, Google DeepMind will slowdown the pace of their research publications. Everyone understands this as a way to protect their “cutting-edge” ideas, and create a competitive advantage with them. DeepMind slows down research releases in battle to keep competitive edge
In any case, some components in the architecture will be open. The acceptance by OpenAI to use Anthropic’s standard to connect the models to data sources within companies is a good example of this. MCP: The new “USB-C for AI” that’s bringing fierce rivals together
Google actually looks worried about this AI performance. They’re re-organizing their AI leadership team. Google Gemini is shaking up its AI leadership ranks
OpenAI finally closed its $40bn funding round. SoftBank is the (generous) investors leading the deal. SoftBank Group Agrees to Lead $40 Billion Funding for OpenAI
AI is creating significant divergences in China’s tech sector. DeepSeek and Alibaba, owners of the most powerful models, are emerging as the winners. Meanwhile, Baidu (which was, in the past, a pioneer in AI) would be among the losers. China’s AI race creates tension at home
AI: Security & Safety
AI safety is turning into a priority, as models become more powerful and more widely adopted
We have confirmation that the current AI models pass the Turing Test. One of the researchers claims that “GPT-4.5 was even judged to be human significantly more often than actual humans”. An AI Model Has Officially Passed the Turing Test
Google DeepMind wants us to anticipate the era of Artificial General Intelligence (AGI), and the associated risks. This week they published a long paper describing their approach to “Technical AGI Safety and Security”. Google says now is the time to plan for AGI safety
The company is optimistic, but calls for responsibility, too. They believe this technology can be a catalyst for progress, but also that its power requires the assessment and prevention of any potential harms that it might cause. Taking a responsible path to AGI
In parallel, Helen Toner, an ex-board member at OpenAI, has published a white paper on this very same topic. She expresses her concern that “we as a society are not ready to handle human-level AI and all its implications”. "Long" timelines to advanced AI have gotten crazy short
Some futurists are seeing a rather gloomy scenario for the coming years. The “AI Futures Project”, a nonprofit led by an ex-OpenAI researcher, based in Berkeley, has released a report and a website called “AI 2027” presenting a rather dark fictional scenario of what could happen with AI in the next 2-3 years. This A.I. Forecast Predicts Storms Ahead
On a related theme, concerns are growing on the future role / purpose of humans, in an “AGI world”:
There is a positive perspective, if we realize that intelligence is not the only factor that makes us human. So rather than trying to compete with AI systems in “raw intelligence”, we should use their capabilities to enhance our own intelligence and promote the other, uniquely human aspects. Being Human in the Age of AI: On Critical Thinking, Agency, and Scientific Discovery
But for now, we don’t seem to be paying attention to the “human factors”. The “Studio Ghibli”-mania with OpenAI, last week, would reveal that we tend to skip the actual “human factors” in art and other human works. Anime lessons in the limits of AI
An additional pessimistic view is that “delegating” mental effort to AI systems could make us more “stupid”. So, according to this view, as AI becomes more powerful, people would lose creativity and become less intelligent. Essay | How I Realized AI Was Making Me Stupid—and What I Do Now
Startups keep working to reduce models’ hallucinations. This makes models less reliable and is a key barrier for adoption (affecting many “sensitive” use cases). Oumi, a startup, claims to have a solution (“HallOumi”). AI lie detector: How HallOumi’s open-source approach to hallucination could unlock enterprise AI adoption
AI-powered surveillance remains a big concern. Even (supposed to be) libertarian states, like Texas, are working hard to deploy these technologies. Texas’ AI-Powered Surveillance Arsenal Has Ballooned. Proposed Laws Provide Few Guardrails.
The next wave of cyberattacks will be AI-powered. AI agents will reduce “barriers to entry” for planning and launching attacks (just like what’s happening with coding…) Cyberattacks by AI agents are coming
AI: Infrastructure
The reports of a slowdown in demand for AI infrastructure might have been greatly exaggerated:
The “Studio Ghibli” storm created an infrastructure bottleneck for OpenAI. This week Sam Altman went so far as to use social networks to call for help. Sam Altman (@sama) on X
Coreweave’s stock rallied this week and is now above the IPO price. Coreweave, the AI infrastructure company that had a poor IPO performance on Friday last week, has now fully recovered from that. CoreWeave’s Two-Day 65% Pop Helps Investors Forget Tepid IPO
AI: Chips
TSMC is under pressure. On one hand, Trump’s ability to set tariffs at light speed are a threat on the company’s export-based business model. On the other hand, the way the company is hedging that risk (a large investment in US facilities) could be too costly. TSMC buys time with $100bn US investment pledge
Intel’s CEO is working on the company’s turnaround… An initial idea he’s mentioned is the spin-off of some non-core businesses. Intel’s New CEO Plots Turnaround; ‘We Need to Improve’
… But it is not clear that investors will give him time. Initial reactions to the CEO’s first messages show skepticism and an impression that he’s proposing nothing too different from what the previous CEO tried. The Clock Is Ticking Loudly for Intel’s New Boss
Quantum Computing
A new scientific result could help build more efficient quantum computers. Yes, we still have to build practical, functioning quantum computers (and it will probably take much longer than some people expect…) but “pseudorandom unitaries” (PRUs), recently discovered by researchers, could make certain quantum algorithms more efficient (and cheaper) to run. The High Cost of Quantum Randomness Is Dropping
Intelligence Augmentation
Augmented Reality
Meta plans to introduce its first smart glasses with a screen by the end of this year. They will have a price tag of over $1,000 and features like hand-gesture controls and a screen for displaying photos and apps. How Meta’s Upcoming $1,000+ Smart Glasses With a Screen Will Work
Brain-Computer Interfaces
Signs of fast progress being made in brain-computer interfaces
25 clinical trials are already under way. Two companies in the US (New York’s based Synchron and Elon Musk’s Neuralink) and one in China (Neuracle Neuroscience) are actively recruiting volunteers to get a brain implant that will connect their brain signals with a computer. Brain-computer interfaces face a critical test
Restoring speech for paralyzed people looks like a killer app. A research team at UC Berkeley is using AI to solve the “latency problem” and make it possible for translate brain signals into audible speech in real time. Brain-to-voice neuroprosthesis restores naturalistic speech - Berkeley Engineering
Neuromorphic computing
Next-gen neural networks could be more similar to biological ones. A research team in Germany has built a next-gen neural network based on “infomorphic neurons”, which mimic biological neurons more closely, and are able to learn with no external coordination (unlike current artificial neural networks). This opens the door to more energy-efficient AI systems, among other things. Self-organizing 'infomorphic neurons' can learn independently
3 - Economic / Business trends
Tech & Geopolitics
The current US push against TikTok could be seen as part of the AI race. ByteDance generates cash from exposure to US consumers and uses it to fund AI technologies in China. AI race gives Washington another reason to be tough on TikTok
It is increasingly clear that US chip bans on China are ineffective. This articles describes the variety of methods that Chinese tech companies are using to bypass these bans. This is how China skirts U.S. chip bans