Planet JM

Tag: llms

  • Open Source is About to Undergo Substantial Change

    …And Most Open Source Communities Aren’t Ready

    It’s probably gauche to talk about “AI” by now. AI this… AI that… and most of the time, what we’re really talking about is predictive text machines, aka LLMs. But today I want to talk about what I see happening in the open source world, and how I see things changing in the not too distant future, and how much of that will be shaped by these predictive text machines, aka… LLMs. The agentic world is growing very quickly, and even if the large LLMs are starting to plateau, the LLM-backed services are still accelerating in their product growth for the simple reason that developers are figuring out how to add rules engines and orchestration platforms to build out targeted vertical services (think tools for reading radiology and MRI scans, for example). A great analogy from computing history for this shift from LLMs to agentic “SLMs” is the shift in emphasis from the single CPU for defining compute power to the emergence of multi-core CPUs along with faster RAM, NVMe, larger onboard caches, and of course, GPUs. When we think about compute power today, we don’t refer to the chip speed, which is a far cry from the late 90’s and early 2000s. Believe it or not, kids, there was a time when many people thought that Moore’s law applied to the clock speed on a CPU.

    For some time now, source code has been of little value. There’s so much of it. Nobody buys source code. I’ve made this point before in a series of posts on the subject. 20 years ago, I noted how internet collaboration was driving down the price of software because of the ubiquity of source code and the ability to collaborate beyond geographic borders. This trend, which has been unceasing now for 25+ years, has hit an inflection point and accelerating beyond the previous rate. This is, of course, because of the oncoming train that is AI, or more specifically, agentic LLM-based systems that are starting to write more and more of our source code. Before I get into the full ramifications of What This Means for Open Source (tm) let me review the 2 previous transformative eras in tech that played a pivotal role in bringing us to this point: open source and cloud.

    Open Source Accelerated the Speed of Development

    A long, long time ago, software vendors had long release cycles, and customers had no choice but to wait 1-2 years, or longer depending on the industry, for the long cycle of dev, test, and release to complete. And then a funny thing happened: more people got online and suddenly created a flurry of core tools, libraries, and systems that gave application developers the ultimate freedom to create whatever they wanted without interference from gate-keepers. I cannot over-emphasize the impact this had on software vendors. At first, it involved a tradeoff: vendors were happy to use the free tools and development platforms, because they saw a way to gain a market edge and deliver faster. At the same time, startups also saw an opportunity to capitalize on this development and quickly create companies that could compete with incumbents. In the late 90s, this meant grabbing as much cash as possible from investors in the hopes of having an IPO. All of this meant that for every advance software vendors embraced from the open source world, they were also effectively writing checks that future competitors would cash, which required that established vendors release even more quickly, lather, rinse, repeat, and find vertical markets where they could build moats.

    Cloud accelerated the speed of delivery

    If open source accelerated the speed of development, the emergence of what became “cloud technologies” enabled the delivery of software at a speed and scale previously thought to be impossible. Several smart companies in the mid-2000s saw this development and started to enact plans that would capitalize on the trend to outsource computing infrastructure. The companies most famous for leading the charge were Amazon, which created AWS in 2006, Netflix, which embraced AWS at an early stage, Google, which created Borg, the predecessor to Kubernetes, and Salesforce, which created it’s cloud-based PaaS, Force.com, in 2009. Where open source gave small growing companies a chance to compete, cloud did the same, but also at a price. Established software vendors started moving to cloud-based systems that allowed them to deliver solutions to customers more quickly, and startups embraced cloud because they could avoid capital expenditures for data center maintenance. Concurrently, open source software continued to develop at a fast pace for the simple reason that it enabled the fast development of technologies that powered cloud delivery. Similar to open source, the emergence of cloud led directly to faster release cycles and increasing competition. Unlike open source, however, cloud computing allowed established cloud companies to build out hegemonic systems designed to exact higher rental fees over time, pulling customers deeper into dependencies that are increasingly difficult to unravel. Software vendors that thought open source developers were the architects of their demise in the early 2000s hadn’t yet met Amazon.

    All of these developments and faster release cycles led to a lot more source code being written and shared, with GitHub.com emerging as the preferred source code management system for open source communities. (Pour one out for Sourceforge.net, which should have captured this market but didn’t.) Sometimes this led companies to think that maybe their business wasn’t cut out for this world of source code sharing, so they began a retrenchment from their open source commitments. I predicted that this retrenchment would have little impact on their viability as a business, and I was right. If only they had asked me, but I digress…

    All of this brings us to our present moment where source code is less valuable than ever. And in a world of deprectiating value for something, how do we ensure that the rules of engagement remain fair for all parties?

    Sorry Doubters: AI Will Change Everything

    If open source accelerated development and cloud accelerated delivery, then AI is accelerating both, simultaneously. Code generation tools are accelerating the total growth of source code; code generation tools are accelerating the ongoing trend of blending the boundary between hardware and software; and code generation tools are (potentially) creating automated systems that deliver solutions more quickly. That last one has not yet been realized, but with the continuing growth of agentic workflows, orchestrators, and rules engines, I would bet my last investment dollar on that trend realizing its potential sooner rather than later.

    What does this portend? I think it means we will need to craft new methods of managing and governing all of this source code. I think it means that rules of collaboration are going to change to reflect shifting definitions of openness and fairness in collaboration. I think it means that previously staid industries (read: semiconductors) are facing increasing pressure in the form of power consumption. speed of data flow, and increasingly virtualized capabilities that have always lived close to the silicon. And I think a whole lot of SaaS and cloud native vendors are about to understand what it means to lose your “moat”. The rise of agentic systems is going to push new boundaries and flip entire industries on their heads. But for the purpose of this essay, I’m going to focus on what it means for rules of collaboration.

    What is the Definition of Open Source?

    For many years, the definition of open source has been housed and governed by the Open Source Initiative (OSI). Written in the post-cold war era of open borders and free trade, it’s a document very much of its time. In the intervening years, much has happened. Open source proliferation happened, and many licenses were approved by the OSI as meeting the requirements of the Open Source Definition (OSD). State-sponsored malware has happened, sometimes inflicting damage on the perceived safety of open source software. Cloud happened, and many open source projects were used in the creation of “cloud-native” technologies. And now LLM-based agentic systems are happening. I mention all of this to ask, in what context is it appropriate to consider changes in the OSI?

    One of the reasons open source governance proved to be so popular is that it paved the way for innovation. Allow me to quote my own definition of innovation:

    Innovation cannot be sought out and achieved. It’s like happiness. It has to be achieved by laying the foundation and establishing the rules that enable it to flourish.

    In open source communities and ecosystems, every stakeholder has a seat at the table, whether they are individuals, companies, governments, or any other body with a vested interest. That is the secret of its success. When you read the 10 tenets of the OSD, it boils down to “Establishing the rules of collaboration that ensure fairness for all participants.” Basically, it’s about establishing and defending the rights of stakeholders, namely the ability to modify and distribute derivative works. In the traditional world of source code, this is pretty straightforward. Software is distributed. Software has a license. Users are held to the requirements of that license. We already saw the first cracks in this system when cloud computing emerged, because the act of distributing… sorry “conveying” software changed significantly when I used software distributed over a network. And the idea of derivative works was formed at a time when software was compiled with shared library binaries (.so and .dll) that were pulled directly into a software build. Those ideas have become more quaint over time, and the original ideas of the OSD have become increasingly exploitable over the years. What use is a software license when we don’t technically “use software”? We chose to not deal with this issue by pretending that it hadn’t changed. For the most part, open source continued to flourish, and more open source projects continued to fuel the cloud computing industry.

    But now we’re bracing for another change. How do we govern software when we can’t even know if it was written by humans? Agentic systems can now modify and write new source code with little human intervention. I will not comment on whether this is a good idea, merely that it is happening. Agentic systems can take the output of cloud-based services, and write entire applications that mimic their entire feature set. Does that meet the definition of open source? Does it violate the EULA of a cloud service? And if companies can recreate entire code bases of projects based only on the requirements of applications that use it, does that violate the terms of reciprocal licenses like the GPL? And this is before we even get to the issues of copyright pertaining to all the source code that had to feed the models in order to write code.

    If we true back to answering the question “how do we protect the rights and ensure the fairness of all participants”, how do we prepare for these changes? I think a couple of things are in order:

    • The right to reverse engineer must be protected to meet the definition of Open Source. This means that the ability to recreate, modify, and redistribute a model, cloud service, or really anything in technology that we use, has to be protected. For years, cloud providers have built in complexity in their services that makes them very difficult to replicate at scale. That is now changing, and it is a good thing.
    • This also means that the ability to recreate, modify, and redistribute models must also be protected if it uses the moniker of Open Source.
    • Agents must abide by licensing terms in order to be categorized as open source. If you call your agentic systems open source, they must be able to interpret and abide by software licenses. This effectively means that all agentic systems will need to include a compliance persona in order to meet the definition of Open Source.
    • Maintainers of Open Source projects must have a way to quickly dismiss the output of agentic systems that file bug and vulnerability reports. This means that in order to meet the open source definition, agentic systems that fit in that category will have to abide by a standard that maintainers use to signal their willingness to accept input from agents. If maintainers decline, then agentic systems will either avoid these projects, or push their inputs and changes into forked repos maintained elsewhere.

    These are just a couple of ideas. The bottom line is that the open source ethos guarantees all stakeholders a seat at the table, and we must be willing to make changes to our governing rules in order to ensure fairness for all parties. To do otherwise is to shirk our responsibility and pretend like it’s still 1999. No change to the open source definition should be taken lightly, but as the governing document that protects the rights of those who participate in open source communities, we need to make sure that it doesn’t become more easily exploitable by monopolistic companies and those that wish to extort from community members or commit harmful acts.

    Open Source communities and maintainers are not yet prepared for these changes, and it’s our job as community members to make sure that these communities, the backbone of open source innovation, remain vibrant and strong.

  • AI Native and the Open Source Supply Chain

    AI Native and the Open Source Supply Chain

    I recently wrote 2 essays on the subject of AI Native Automation over on the AINT blog. The gist of them is simple:

    It’s that latter point that I want to dive a bit deeper into here, but first a disclaimer:

    We have no idea what the ultimate impact of "AI" is to the world, but there are some profoundly negative ramifications that we can see today: misinformation, bigotry and bias at scale, deep fakes, rampant surveillance, obliteration of privacy, increasing carbon pollution, destruction of water reservoirs, etc. etc. It would be irresponsible not to mention this in any article about what we call today "AI". Please familiarize yourself with DAIR and it's founder, Dr. Timnit Gebru.

    When I wrote that open source ecosystems and InnerSource rules were about to become more important than ever, I meant that as a warning, not a celebration. If we want a positive outcome, we’ll have to make sure that our various code-writing agents and models subscribe to various agreed-upon rules of engagement. The good news is we now have over 25 years of practice for open source projects at scale that gives us the basis to police whatever is about come next. The bad news is that open source maintainers are already overwhelmed as it is, and they will need some serious help to address what is going to be an onslaught of “slop”. This means that 3rd party mediators will need to step up their game to help maintainers, which is a blessing and a curse. I’m glad that we have large organizations in the world to help with the non-coding aspects of legal protections, licensing, and project management. But I’m also wary of large multi-national tech companies wielding even more power over something as critical to the functioning of society as global software infrastructure.

    We already see stressors from the proliferation of code bots today: too many incoming contributions that are – to be frank – of dubious quality; new malware vectors such as “slopsquatting“; malicious data injections that turn bots into zombie bad actors; malicious bots that probe code repos for opportunities to slip in backdoors; etc – it’s an endless list, and we don’t yet even know the extent to which state-sponsored actors are going to use these new technologies to engage in malicious activity. It is a scary emerging world. On one hand, I look forward to seeing what AI Native automation can accomplish. But on the other, we don’t quite understand the game we’re now playing.

    Here are all the ways that we are ill prepared for the brave new world of AI Native:

    • Code repositories can be created, hosted, and forked by bots with no means to determine provenance
    • Artifact repositories can have new projects created by bots with software available for download before anyone knows no humans are in the loop
    • Even legitimate projects that use models are vulnerable to malicious data injections, with no reliable way to prove data origins
    • CVEs can now be created by bots, inundating projects with a multitude of false positives that can only be determined by time-consuming manual checks
    • Or, perhaps the CVE reports are legitimate, and now bots scanning for new ones can immediately find a way to exploit one (or many) of them and inject malware into an unsuspecting project

    The list goes on… I fear we’ve only scratched the surface of what lies ahead. The only way we can combat this is through the community engagement powers that we’ve built over the past 25-30 years. Some rules and behaviors will need to change, but communities have a remarkable ability to adapt, and that’s what is required. I can think of a few things that will limit the damage:

    • Public key architecture and key signing: public key signing has been around for a long time, but we still don’t have enough developers who are serious about it. We need to get very serious very quickly about the provenance of every actor in every engagement. Contributed patches can only come from someone with a verified key. Projects on package repositories can only be trusted if posted by a verified user via their public keys. Major repositories have started to do some of this, but they need to get much more aggressive about enforcing it. /me sideeyes GitHub and PyPi
    • Signed artifacts: similar to the above – every software artifact and package must have a verified signature to prove its provenance, else you should never ever use it. If implemented correctly, a verified package on pypi.org will have 2 ways to verify its authenticity: the key of the person posting it, and the signature of the artifact itself.
    • Recognize national borders: I know many folks in various open source communities don’t want to hear this, but the fact is that code that emanates from rogue states cannot be trusted. I don’t care if your best friend in Russia has been the most prolific member of your software project. You have no way of knowing if they have been compromised or blackmailed. Sorry, they cannot have write access. We can no longer ignore international politics when we “join us now and share the software”. You will not be free, hackers. I have to applaud the actions of The Linux Foundation and their legal chief, Michael Dolan. I believe this was true even before the age of AI slop, but the emergence of AI Native technologies makes it that much more critical.
    • Trust no one, Mulder: And finally, if you have a habit of pulling artifacts directly from the internet in real time for your super automated devops foo, stop that. Now. Like.. you should have already eliminated that practice, but now you really need to stop. If you don’t have a global policy for pushing all downloads through a centralized proxy repository – with the assumption that you’re checking every layer of your downloads – you are asking for trouble from the bot madness.
    • Community powered: It’s not all paranoid, bad stuff. Now is a great opportunity for tech companies, individual developers, enterprises, and software foundations to work out a community protocol that will limit the damage. All of these actors can sign on to a declaration of rules they will follow to limit the damage, quarantine known bad actors, and exchange vital information for the purpose of improving security for everyone. This is an opportunity for The Linux Foundation, Eclipse, and the Open Source Initiative to unite our communities and show some leadership.
    • Bots detecting bots: I was very hesitant to list this one, because I can feel the reactions from some people, but I do believe that we will need bots, agents, and models to help us with threat detection and mitigation.

    I have always believed in the power of communities to take positive actions for the greater good, and now is the perfect time to put that belief to the test. If we’re successful, we can actually enjoy revamped ecosystems that will be improved upon by our AI Native automation platforms. If successful, we will have safer ecosystems that can more easily detect malicious actors. We will also have successful communities that can add new tech capabilities faster than ever. In short, if we adapt appropriately, we can accelerate the innovations that open source communities have already excelled at. In a previous essay, I mentioned how the emergence of cloud computing was both a result of and an accelerant of open source software. The same is true of AI Native automation. It will inject more energy into open source ecosystems and take them places we didn’t know were possible. But what we must never forget is that not all these possibilities are good.