195 lines
14 KiB
Markdown
195 lines
14 KiB
Markdown
# An Open Future
|
||
*Source: https://openfuture.tenstorrent.com/*
|
||
*Version: V1.0 4/2025*
|
||
|
||
## Mapping the Open Territory
|
||
|
||
---
|
||
|
||
## Part 1: How We Got Here
|
||
|
||
AI is changing the laws that once governed computing.
|
||
|
||
Until recently, Bell's Law gave us an accurate frame for understanding computing revolutions, stating that each decade a new class of computing emerges, resulting in a fundamental shift in access¹.
|
||
|
||
We went from mainframes in the 1950s, to minicomputers in the 1960s, to super computers in the 1970s, to personal computers in the 1980s, to the world-wide web in the 1990s, and mobile in the 2000s.
|
||
|
||
These revolutions allowed us to make computers that were much more accessible – simultaneously driving performance up 10X while also driving cost down 10x. In 1981, a fully loaded IBM PC cost $4500². Today, an iPhone, which is many millions of times faster³, retails for $1,129⁴. Through this process we got very good at building very powerful computers with very small chips.
|
||
|
||
### Timeline of Open
|
||
|
||
**THE FIRST COMMERCIAL MAINFRAME COMPUTER, RELEASED IN 1948 BY THE ECKERT-MAUCHLY COMPUTER CORPORATION (EMCC).**
|
||
|
||
*Timeline showing evolution from 1950-2000:*
|
||
- 1950: MAINFRAMES
|
||
- 1960: MINICOMP
|
||
- 1970: PC
|
||
- 1980: BROWSER
|
||
- 1990-2000: Various milestones including UNIVAC, 12-BIT PDP-8 IC CHIP, INTEL 4004, MINITEL, WWW, LINUX, "OPEN SOURCE", MOZILLA, RED HAT, DRAM, IBM ANTI-TRUST LAWSUIT, UNIX, ETHERNET
|
||
|
||
---
|
||
|
||
AI is changing the laws that once governed computing.
|
||
|
||
AI is valuable enough to warrant this kind of investment. It is literally, as Andrej Karpathy said, "Software 2.0"⁸.
|
||
|
||
It isn't just an efficiency gain, like previous revolutions. AI creates knowledge that we didn't have before. It is unprecedented how quickly AI can navigate nearly inconceivable amounts of data and complexity. It will ask questions we didn't even know to ask. It will destroy previous industries and create new ones. Those that know how to leverage it, and can afford to, will reap the rewards.
|
||
|
||
But we can't assume that we'll return to the historical trend of falling costs and broadening access. We're at a critical juncture. As companies build out their AI stack, they are making a choice today that will determine the future. Companies can invest in closed systems, further concentrating leverage in the hands of a few players, or they can retain agency by investing in open systems, which are affordable, transparent, and modifiable.
|
||
|
||
But we can't assume that we'll return to the historical trend of falling costs and broadening access. We're at a critical juncture. As companies build out their AI stack, they are making a choice today that will determine the future.
|
||
|
||
Every shift created new leaders, sidelined old ones, and required adaptation. From a social perspective, these innovations gave many more people access to compute.
|
||
|
||
However, prices aren't dropping with the advent of Artificial Intelligence. While cost per math operation is going down, the actual cost of inference per token is still climbing⁹ as models are getting larger (eg. GPT-4⁵), doing more work (e.g. "reasoning models"), and doing work that is more intensive (e.g. new GPU generation). AI datacenters are orders of magnitude more powerful than previous generations with spending rising by tens of billions year-over-year⁶. Even if we eventually see some cost reductions, it will take time before they reach affordability, leaving everyone besides a few in the dust of the AI revolution.
|
||
|
||
Why is this computer class more expensive? AI is extremely physically intensive – requiring more silicon, more energy, more resources. From shifting the physics of compute at the transistor level to building out the global infrastructure of AI data centers, this revolution is pushing against the physical limitations of human industry⁷.
|
||
|
||
If Bell's Law breaks fully, AI will be the first computing revolution that doesn't increase access, but instead concentrates it. We saw hints of this concentration effect with the previous computer class. Jonathan Zittrain argues that the cloud has put accessibility at risk leaving "new gatekeepers in place, with us and our limited business plans and to regulators who fear things that are new and disruptive⁹." Unlike hyperscalers before it, AI threatens to tip consolidation into full enclosure.
|
||
|
||
If AI eats everything, like software has eaten everything¹⁰, this means that open versus closed is a referendum on the future shape of society as a whole.
|
||
|
||
A handful of companies will own the means of intelligence production, and everyone else will purchase access at whatever price they set. As many have warned, this will represent a new form of social stratification.
|
||
|
||
It is clear to us that open is existential.
|
||
|
||
---
|
||
|
||
## Part 2: A Closed World
|
||
|
||
This isn't the first time we've been presented with a choice between a closed or open future. In fact, we're living in a closed world today because of choices made for us 40+ years ago. Early minicomputer and PC culture was dominated by a hacker ethos defined by "access to computers... and the Hands-On Imperative¹¹." By the late 90s and early 00s, PC development became dominated by Windows and Intel at the cost of limiting innovation while hamstringing¹² competitors and partners alike¹³.
|
||
|
||
How do closed worlds form? One word: swamps. A swamp is a moat gone stagnant from incumbents who have forgotten how to innovate.
|
||
|
||
### Innovation Ownership Diagrams
|
||
|
||
**FIGURE 1. CLOSED**
|
||
- Shows a single "VERTICAL OWNER" in the center
|
||
- No leverage or choice in dealings
|
||
|
||
**FIGURE 2. PROPRIETARY**
|
||
- Shows "PROPRIETARY OWNER" surrounded by multiple "CUSTOMER" boxes
|
||
- No control of roadmap or features while incurring higher development and product costs
|
||
|
||
**FIGURE 3. OPEN**
|
||
- Shows "OPEN FOUNDATION" surrounded by multiple "CUSTOMER" boxes in a collaborative arrangement
|
||
- You drive and control the future
|
||
|
||
The writing is on the wall for AI. We are veering towards a closed world where the constellation of technology companies are fighting over scraps. Competition, innovation, and sustainable business can't thrive in this low-oxygen environment.
|
||
|
||
How do closed worlds form? One word: swamps. A swamp is a moat gone stagnant from incumbents who have forgotten how to innovate.
|
||
|
||
There are many ways to produce a swamp. They can protect a product by overcomplicating it, adding unnecessary proprietary systems and layers of abstraction. They can charge rents, in the form of license fees. They can pile on features just enough to justify an upgrade to customers, while staying disconnected from what they actually need. And if they want to get really clever, they can offer something "for free" as an inseparable part of a bundled service in order to lock out competition.
|
||
|
||
However it happens, what started as innovation becomes just an extra tax on the product, erecting monopolies instead of creating real value. These companies become incentivized to preserve the status quo, rather than changing.
|
||
|
||
But, as we've seen before, the world always changes.
|
||
|
||
---
|
||
|
||
## Part 3: An Open World
|
||
|
||
Open source has a way of infiltrating crucial computing applications. The internet runs on it¹⁹. The entire AI research stack uses open source frameworks. Even proprietary tech relies on it with 90% of Fortune 500 companies using open source software²⁰. There wouldn't be macOS without BSD Unix, Azure without Linux, or Netflix without FFmpeg.
|
||
|
||
Open source and its hardware equivalent, open standards, have repeatedly catalyzed mass adoption by reducing friction and enabling interoperability. Robert Metcalf says the openness of ethernet allowed it to beat rival standards²¹. DRAM enabled the mass adoption of PCs with high-capacity, low-cost memory, while PCIe enabled high-speed interoperability of PC components. Similarly, Open Compute Project specs, used by Meta and Microsoft among others, standardized rack and server design, so components could be modular and vendor-agnostic²².
|
||
|
||
RISC-V is the hardware equivalent of Linux for AI hardware.
|
||
|
||
RISC-V is the hardware equivalent of Linux for AI hardware. It launched in 2010 at UC Berkeley as a free, open standard alternative to proprietary architectures like Intel's x86 and ARM²³. Its open nature allows it to be deeply customized, making it especially desirable for AI and edge computing applications, and it is royalty-free. RISC-V's ISA is gaining incredible adoption, with companies from Google to us at Tenstorrent adopting it for custom silicon.
|
||
|
||
Open systems also attract a global talent pool. Linux itself is the shining example of this, constructed by thousands of engineers, with significant contributions coming both from independent outsiders and employees of major players like Intel and Google²⁴.
|
||
|
||
We believe open is the default state – what remains when artificial boundaries fall away. The only question is how long those boundaries hold, and how much progress will be delayed in the meantime.
|
||
|
||
### The AI Stack - Closed Today
|
||
|
||
Today, parts of the AI stack are open, parts are closed, and parts have yet to be decided. Let's look at a few of the layers:
|
||
|
||
#### 🔧 HARDWARE
|
||
**● CLOSED**
|
||
|
||
Most hardware today is a black box, literally. You're reliant on a company to fix, optimize, and, at times, even implement your workloads.
|
||
|
||
#### 📊 LOW LEVEL SOFTWARE
|
||
**● CLOSED**
|
||
|
||
Most parallelization software is proprietary causing unnecessary lock-in and massive switching costs.
|
||
|
||
#### 🧠 MODELS
|
||
**● MIXED**
|
||
|
||
Models are mixed, but most of the leading ones are closed. The models that are open share limited data, with little to no support, and have no promises of staying open in the future.
|
||
|
||
#### </> APPLICATIONS
|
||
**● CLOSED**
|
||
|
||
Even if an application is using an open source model, most are built using cloud platform APIs. This means your data is being pooled to train the next gen models.
|
||
|
||
The current stack tells a story of closed engulfing open, stopping innovation in its tracks – a classic swamp.
|
||
|
||
Opening up AI hardware, with open standards like RISC-V, and its associated software would trigger a domino effect upstream. It would enable "a world where mainstream technology can be influenced, even revolutionized, out of left field²⁵." This means a richer future with more experimentation and more breakthroughs we can barely imagine today, like personalized cancer vaccines²⁶, natural disaster prediction²⁷, and abundant energy²⁸. And this world gets here a lot faster outside of a swamp.
|
||
|
||
There's an old Silicon Valley adage – if you aren't paying you are the product. In AI, we've been paying steeply for the product, but we still are the product. We have collectively generated the information being used to train AI, and are feeding it more every day.
|
||
|
||
In a closed world, AI owns everything, and that AI is owned by a few. Opening up hardware and software means a future where AI doesn't own you.
|
||
|
||
---
|
||
|
||
## Part 4: Building an Open Future
|
||
|
||
At Tenstorrent, we're committed to building an open future for AI.
|
||
|
||
Open can mean a lot of things. For us, open means affordable, transparent, and modifiable.
|
||
|
||
### AFFORDABLE
|
||
|
||
AI hardware shouldn't be a luxury product. Universal access to intelligence requires reasonable costs. The future deserves a proliferation of AI applications, not just a few businesses capable of surviving on tiny margins thinned by monopoly rents.
|
||
|
||
### TRANSPARENT
|
||
|
||
You don't really own it unless you understand what you own, which is why we don't sell black boxes. Our hardware is built on open standards, with each layer of the stack built from first principles for complete navigability resulting in transparency.
|
||
|
||
### MODIFIABLE
|
||
|
||
You should be able to choose what you want and what you don't want. Open shouldn't be another form of control. It should empower you to create your own tech stack that suits your specific needs.
|
||
|
||
It will not be easy to achieve this open future. Hardware resists openness, and software isn't exempt either. Most developers rely on copyright law, which is automatic and offers the same protection for jingles and songs. Change a few lines of code, or a shape in a drawing, and it's a new work. Software patents muddy the waters, locking down broad concepts with vague claims. And hardware's worse where patents are the default. Surmounting the burden of patent law means we need to create a full-stack hardware and software company, or create a consortium of companies.
|
||
|
||
We, at Tenstorrent, are doing both.
|
||
|
||
To that end, we're building up organizational excellence across multiple verticals from hardware to software because if we don't, then closed systems will continue to block innovation. It's necessary that the entire stack be open, otherwise we'll remain in the swamp we're in today.
|
||
|
||
We are also opening up our technology. Our IP is transparent, our architectures are open, and our software is open source so you can edit, select, fork, and own your silicon future.
|
||
|
||
Join us.
|
||
|
||
## Own Your Silicon Future
|
||
|
||
[Tenstorrent Github →](https://github.com/tenstorrent)
|
||
|
||
---
|
||
|
||
*References:*
|
||
1. Bell's Law reference
|
||
2. IBM PC cost reference
|
||
3. iPhone performance comparison
|
||
4. iPhone pricing reference
|
||
5. GPT-4 reference
|
||
6. AI datacenter spending reference
|
||
7. Physical limitations reference
|
||
8. Andrej Karpathy "Software 2.0" reference
|
||
9. Cost per token reference
|
||
10. "Software eating everything" reference
|
||
11. Hacker ethos reference
|
||
12. Innovation limitation reference
|
||
13. Competitor hamstringing reference
|
||
19. Internet open source reference
|
||
20. Fortune 500 open source usage
|
||
21. Robert Metcalf ethernet reference
|
||
22. Open Compute Project reference
|
||
23. RISC-V Berkeley reference
|
||
24. Linux global talent reference
|
||
25. Technology influence reference
|
||
26. Personalized cancer vaccines reference
|
||
27. Natural disaster prediction reference
|
||
28. Abundant energy reference |