🔥RTN: How Facebook came to be a leader in open-source
Fervo Energy, Apple 3D printing, Europe's Industrials are in trouble
💡 If you enjoy this newsletter, please forward it to your smartest friend 🙏🏻
Facebook this week released an open-source Gen AI model LlaMa 2, in partnership with Microsoft. You know, I’m not here to talk about Gen AI though.
What’s more interesting is that Meta has been shipping a lot of amazing open-source software of late, across a number of domains like robotics, simulations and compute.
Why’s that?
In 2021 Meta’s pivot to the metaverse was met with a yawn. The company has also been under Regulatory scrutiny, with accusations of monopolistic rent extraction and the undermining of democracy using social (remember that!).
Their old strategy of ‘closed’ is being rapidly replaced with a new strategy of ‘open’, with Meta placing open-source products at the centre of multiple fields. They’re unlikely to be able to do any big acquisitions and aligning openly with partners (like Microsoft) and developers is prudent. Being ‘open’ is a foil to the regulatory heat, and is strategically smart.
So, let’s take a (non-exhaustive) look at what they’ve been doing in our part of the tech world;
Semiconductors and PyTorch: Nvidia has a monopoly on AI workloads, largely due to CUDA, a parallel computing program for GPUs. To date, AI software development has had to be built on CUDA. Meta is displacing this, with the PyTorch framework which is seeing increasing developer adoption. It runs on Python (the dominant ML language), is Open Source and allows computation on a wider variety of chips, be it AMD, ARM, ASICs etc. I’m personally bullish on AMD due to this. (OpenAI’s Triton is also displacing CUDA, worth reading into)
Digital Twins (more below): this week, Meta’s Reality Labs released an open dataset, Aria. An ‘egocentric’ dataset designed for ‘Digital Twins’, a hyper-accurate dynamic digital counterpart for scene replication. Meta uses proprietary glasses fitted with a number of cameras to collect data to render the scene. Though its primary use is in AR/VR but has applications in multi-modal learning, robotics, simulation, predictive maintenance etc.
Computer Vision: Meta released their open-source Segment Anything Model (SAM) earlier this year. Improving on the hard CV problem of image segmentation. SAM can generate segmentation masks for objects in images, either through general detection or providing points for specific objects. We’re likely to see a number of improved applications in robotics using SAM
winning over the Hacker News crowd ain't easy
⚡️Energy, Materials and Climate
“the most productive enhanced geothermal project ever completed” Fervo’s CEO
Well, well, well. Fervo Energy successfully tests geothermal well in Texas (paywall)
Why it matters: Enhanced Geothermal Systems (EGS) is a rare new entrant to the renewable mix. EGS is interesting as its less geographically constrained, can be used to store energy and can produce 24/7 zero-carbon power
This week, Fervo announced the results of a 30-day EGS connection test in Nevada. They produced enough hot water to generate about 3.5 megawatts of electricity (enough to power 2,500 homes)
Geothermal involves tapping into high-temperature geothermal reservoirs, located in the Earth's crust. Water / steam is pumped to the surface and used to turn turbines connected to generators, producing electricity
Fervo Energy's approach to geothermal enables in-reservoir energy storage and dispatchable generation attributes. The technology provides sustained heat recovery and energy storage capabilities with high round-trip efficiencies, making geothermal a reliable resource for generating low-carbon electricity
Unlike nuclear fission, which is regulated to near-oblivion, geothermal faces relatively fewer policy obstacles. Drilling on federal lands involves federal permitting, similar to fracking
However, the current levelised cost of energy (LCOE) is around $450/MWh, which as you can see has some way to being competitive within the renewable mix:
IEA’s Electricity Report Update = European Industrial in trouble
Why it matters: This threat to Europe’s industrial capacity, is an opening for founders building new, technological solutions. These industries have historically sat comfortably due to CAPEX moats, they now have multiple headwinds across cost inputs (energy+labour) and trade geopolitics.
🚖 Moving things
Good thread on the FAA’s new proposed regulations for eVTOLs in the US
Why it matters: progress in deploying new aircraft has been stymied by not only power to weight ratio of batteries but largely by unclear regulation and infrastructure. This new bill is a shot in the arm for aircraft companies looking to build a new travel paradigm
Summary:
Using the same regulatory pathways as existing aircraft
No special verti-ports to be used, aircraft will use airports and heliports
UTM is out, manual air traffic control operators are to be used
🦾 Manufacturing and Robotics
Apple is rumoured to be soon using 3D printing for the Apple Watch
Why it matters: Additive Manufacturing has been stuck in a low-volume production trend line for some time. Apple using AM for mass production in their watches might lower costs and broaden the use of the technology, though currently this is being lead by China
Apple is rumoured to be using AM for mechanical components its Ultra Watch
China, once again, is ahead in the use of this technology at scale
Though, there are exceptions like OECHSLER in polymers, Oerlikon in metals and a number of space companies leaning hard into AM
Should Apple successfully integrate 3D printing into its serial and mass production processes for the Apple Watch Ultra, it is likely to expand its use of this technology across other products