Everything that happened in deeptech in 2024
What a year! I donāt know about you, but Iām looking forward to a few days off to unwind from a crazy year. I wanted to provide a quick write-up on some of the most interesting and notable stories, trends and technological advances of 2024. Some digestible deeptech tidbits to share with your loved ones over the holiday period š¤. More below.
š£ Announcements
š§š¼āāļø Taking a break: I wonāt be writing a monthly update at the start of January, as Iām having a much-needed week off. Iāll be back at the start of February, so donāt despair :)
š Annual round-ups. Last year, I wrote a series of 2023 roundups across āRobotics and Manufacturingā, āClimateā and āGeopolitics and Defenceā. Hit reply if youād like to see the same for 2024
š¬2024 events
We hosted our annualĀ REALTECH ConferenceĀ in September, overlooking sunny London (full coverageĀ here). We also held community dinners in Milan, Zurich, London, and Berlin, with many of you in attendance. Next year,Ā we want your help growing the community further.Ā Expect some announcements early next year about a full program of events that will further accelerate progress in our fields of interest. If youād like to co-host an event in your city, hit me up!
š¤ Research
Alongside Dealroom and launched at the REALTECH Conference, we wrote the definitive report on European REALTECH
We compiled 40+ pages of in-depth analysis on trends in REALTECH. We analysed 15k companies, 11k investors over 10 years - read it here
Now for everything else that went down this yearā¦..
Robotics and Manufacturing
šŗļøāChatGPT for the 3D worldā - OpenAI released their first multi-modal model 4o in May, combining language and vision. Since then, the race to build general-purpose models for the physical world has intensified. Numerous vaunted computer vision academics have set out to build their own companies. With notable rounds such as Fei-fei Liās World Labs ($230m), Skild AI ($300m), Physical Intelligence ($70m), and ArchetypeAI ($13m).
šļø Connecting language to vision:Ā Applied research in Vision-Language Models (VLM) and Vision-Language-Action (VLA) continued to gain ground. These models combine internet-scale textual data with vision data.Ā WayveĀ released LINGO-2, the first VLA capable of autonomously driving cars using a natural language interface.Ā Physical IntelligenceĀ released theĀ Ļ0Ā model,Ā focusing on high-frequency dexterity tasks (e.g., folding laundry or bussing tables). It adapts internet-scale semantic knowledge to low-level motor control. However, some of these multi-modal approaches are shown to be dead ends.Ā
š Physics meets deep learning - there has been more work integrating and understanding physics into machine learning models. DL models are good at learning complex non-linear patterns, many of which are found in physics. Notably, ArchetypeAIās Newton learns foundational physics principles from 600 million cross-modal sensor measurements, enabling zero-shot inference and accurate predictions across diverse systems. PhysicsXās Large Physics Model (LPM) which is being used for generative design engineering in aerospace. Lastly, Universal Physics Transformers (UPT) embeds physics into its architecture by creating a latent space representation of complex simulations, allowing for efficient prediction and simulation of physical systems without relying on granular data.
š¤ Robotics goes open-sourceāmany frameworks, datasets, and models were open-sourced to accelerate adoption and progress for robotics practitioners. Notable contributions includeĀ The Well,Ā a 15-terabyte collection of datasets containing numerical simulations of various spatiotemporal physical systems, and an updatedĀ ALOHA-2, low-cost hardware for bimanual teleoperation.Ā š¤ Hugging FaceĀ capitalised on more developers building in robotics and launched theĀ LeRobotĀ robotics framework.Ā
āWe're at the beginning of a new industrial revolution. The next wave of AI is physical AI. AI that understands the laws of physics, AI that can work among usā Jensen Huang, CEO of NVIDIA
š¢ NVIDIA doubles down on the āNext Industrial Revolutionā - NVIDIA doubled down on technologies at various levels of the robotic autonomy stack. Notably, updates to Isaac Sim and Isaac Manipulator - a simulation platform and workflow for robotic arms, respectively. They also launched the Generalist Embodied Agent Research (GEAR) Lab, led by Dr Jim Fan, a research lab working on embodied AI research, releasing work such as Project Gr00t, a foundational humanoid robotics model.
š Autonomous vehiclesāanother seminal year as winners were anointed, and the also-rans bow out (dis)gracefully. Wayve raised a flurry of leading applied research from PRISM-1 to WayveScenes and raised $1bn from Softbank. Their end-to-end (E2E) learning approach has been deemed the winner, with Tesla pivoting its architecture to E2E with their release of FSD v12. Uber returned to AVs, with partnerships with Wayve and Waymo. The deals will likely enable autonomy companies to run their own autonomous fleets on the app, leveraging Uberās distribution advantage.
š¹ AppleĀ finally cannedĀ Project Titan, its secretive electric and autonomous vehicle project, started in 2014.Ā General MotorsĀ threw in the towel onĀ Cruise;Ā after spending $10 bn, it has given up on the possibility of launching robotaxis, citing āan increasingly competitive robotaxi marketā and āconsiderable time and resources.āĀ Innovatorās dilemma, much?
š¤ Humanoid demos are everywhere.Ā Humanoid makers made the most of the peak hype cycle by releasing a flurry of flashy demonstrations. Agility Robotics showed off itsĀ Digit humanoid working within a GXO logistics warehouse, andĀ Boston Dynamics released its new humanoid robot, Atlas (video above). Though, Humanoid deployment in commercial settings at scale remains some way off.
ā°ļø Legacy automakers: āWe aināt dead yet!ā - another torrid year for automakers struggling to transition from internal combustion engines to electric vehicles. European automakers such as VW, Mercedes and Stellantis issued profit warnings as low-cost Chinese EVs continue to flood the European market. Symptomatic of the EV challenge, VW partnered with EV company Rivian in a $5.8bn joint venture to license Rivianās software-defined infrastructure. VW gets best-of-breed EV software architecture, likely putting the nail in the coffin of VWās internal software unit Cariad.
Defence and Geopolitics
š”ļø AI in defence - AI has become a critical new domain for defence departments, as they look to harness the technology for faster decisions and broader capabilities. Private sector firms have stepped in to integrate their capabilities, launching more national security-specific models. Such as Scale AIās āDefense Llamaā and Palantir & Anthropicās āClaude 3.5ā for defence. Working alongside departments - such as the US Armyās Project Linchpin, the first program of record to integrate third-party AI models into their forces. However, critical questions around ethics, safety and oversight have hampered adoption. The UKās MoD recently published the āDependable AI in Defenceā directive to help address these questions. Many current deployments maintain āhumans-in-the-loopā and focus on upgrading legacy systems versus net new autonomous capabilities.
š Drones - moved into the autonomy era with vendors building new capabilities for security. These include GPS-denied navigation, swarming and low-cost, attritable units. The approaches to GPS-denied navigation and autonomy are interesting, with various approaches such as inertial sensor navigation, visual localisation and even quantum geomagnetic wayfinding. We saw notable product releases such as Palantirās Visual Navigation (VNav) and Helsingās HX-2 drone which uses a third-party attritable low-cost drone enabled with their own software.
Palantir is the first software prime - the company has shipped a multitude of AI systems to the Department of Defense (DoD), such as TITAN ($178m - autonomous target recognition), MAVEN ($480m - sensor fusion system), and just this month, their drone visual navigation technology (VNav). They now surpass other traditional primes in terms of market capitalisation:
AndurilĀ is also prime now.Ā In the development-for-production phase of theĀ collaborative combat aircraftĀ (CCA)Ā drone program, Anduril and General Atomics beat out giants Lockheed Martin, Northrop Grumman, and Boeing. This is a marked event, given that the company was founded in 2017.
āļøāš„Technology sanctions donāt work - sanctions imposed mainly by the US on technology exports to China arguably arenāt working. Firstly, China has been evading export bans on key technologies such asĀ Nvidia chips and advanced semiconductor tooling.Ā ChinaāsĀ leading-edge chip manufacturer,Ā SMIC, is thought to be three years behind TaiwanāsĀ TSMC. Additionally,Ā ChinaāsĀ leading LLM models are at least at par with Western equivalents.Ā China TelecomĀ is rumoured to haveĀ trained a 1 trillion parameter language model on domestic silicon. Technology sanctions may be accelerating Chinaās development of its own state capacity.
Climate Tech
ā”ļø Big Tech revolutionises clean energy financing - clean energy has become the critical bottleneck to AI datacentre buildout needed to scale AI computing. Big energy projects and meet growing demand. Googleās Clean Transition Tariff (CTT) with NV Energy supports Fervo Energyās 115 MW geothermal project for Nevada data centres, while Amazon, Google, and Microsoft partnered with Duke Energy on Accelerating Clean Energy (ACE) tariffs for emerging clean technologies.
ā¢ļø Big Tech goes Nuclear on AI and forgets about Net Zero goals - Though as clean baseload power is a limited resource, Big Tech has turned to Nuclear power. Google partnered with Kairos Power for up to 500MW of power, Amazon announced a partnership with X-Energy for up to 960MW, and Microsoft reached a deal with Constellation Energy to restart nuclear power on Three Mile Island, home of one of only three nuclear meltdowns. In doing so, they have quietly rolled back their Net Zero targets.
š° From āFOAKā to āNOAKā climate financing - we saw the first projects in the US go from āfirst of a kindā financing to ānext of a kindā financing. Namely, Solugenās $214m loan guarantee from the US Department of Energy (DOE) Loan Programs Office (LPO) to build their second commercial plant in Minnesota, a āBioForgeā facility to produce bio-based organic acids. These financings have been made possible due to Bidenās Inflation Reduction Act, now under threat from the Trump presidency.
š¬ Progress in materials discovery - Transformer architecture has reduced search complexity in non-deterministic fields such as biology and chemistry. Weāve seen numerous leaps in progress this last year. Metaās Fundamental AI Research (FAIR) and Georgia Tech released OpenDAC, a dataset and challenge to find new sorbents for Direct Air Capture (DAC) using computational chemistry. Solugen used Bayesian optimisation to discover new chemicals. Orbital Materials, an AI materials discovery startup founded by an ex-DeepMind researcher, has open-sourced its materials simulation model, Orb. We also saw LLMs used to build autonomous lab workflows, such as LLMatDesign - an autonomous materials discovery agent.
šŖšŗ Northvolt and the EUāThe European Union has passed many regulations to force market participants toward Net Zero. However, it has not developed a working model for which renewable industries to support, especially those in the cross hairs of Chinese low-cost producers. The EU has decided to support solar cell manufacturers with ā¬400bn, competing against low-cost producers in China. Though chose not to do so, for now-bankrupt Northvolt, the European sovereign battery manufacturer that raised $15bn. With Northvoltās demise, Chinese battery producer CATL wants to build a ā¬4bn battery manufacturing facility in Spain.
āļø The West is getting serious about critical minerals. The EU enacted the Critical Raw Materials Act (CRMA), which seeks to diversify the continent's supply and improve the recycling and circularity of materials. The 34 materials listed are critical to net-zero technologies, such as batteries, solar cells, and wind turbines. China has just banned US exports of critical materials gallium, germanium, and antimony, which are used in military applications.
š New geothermal technology for baseload powerāFervo Energy has proven the consistent energy generation of its Enhanced Geothermal Systems (EGS) in Utah. This breakthrough is a boon for geothermal energy, which can provide 24/7 clean baseload energy from the heat of the earthās crust.
šŗšø Trump in, climate policy out? Trumpās election has triggered fear amongst climate advocates, given his stance as a climate sceptic. Trump has nominated long-time oil and gas executive Chris Wright for secretary of the Department of Energy (DOE). Trump has vowed to roll back all Biden-era climate policies, including the IRA, which has been a huge accelerant to climate projects.
DeepMind predicts the weather. It has released the generative diffusion model GenCast, which builds on its earlier release, GraphCast. GenCast can predict 15-day weather with increased accuracy and significantly lower inference time than leading physics-based weather models (20% better than ENS forecasts). This new capability is much needed, given the increase in extreme global weather events.
Would love to see a robotics report! Always amazing updates :) Enjoy your well-deserved time off!