AI in 2024: The Future is …​

We are well into the surge of Large Language Models (LLMs). Half a year in, it’s prime time to assess their progress and discern any emerging trends. Lo and behold, an AI adoption trajectory is indeed taking clear shape. Let’s look at people making fast AI money today.

Figure 1. DALL-E a typically distorted image edit of the author

The Context of this Analysis

Context: Before diving in, let me set the stage for my observations. Since the 90s, I’m privileged to follow a fluent community known by the moniker The "Northern." These folks gathered in Québec this year hosted by a subgroup of Waterloo alumni. The Northern has been an invite-only assembly of independent, expert programmers who generally operate as Solo Entrepreneurs — a concept described in the book "Company of One: Why Staying Small Is the Next Big Thing for Business" by the esteemed Paul Jarvis, one such person. This community and many like it is made up of loosely related special interest groups referred to as "Cliques". Many members counting over 500 small businesses in 2023 are seasoned software engineering professionals who perhaps found the corporate world not challenging or fast-paced enough. Having learned how to sell and operate independently, these professionals embark on a new journey, honing their skills and embracing the lifestyle of modern Digital Nomads. Freedom rather than greed is the key appealing factor in this demographic. Nature then leads these independent professionals to seek like minds with like interests forming many classic hacker cliques making up the bigger whole. When I first met this community, a smelly single-room BBS office in Philadelphia was its home. Today its Discord.

It is important to note that labeling the Northern as a "group" from demogroup, as I see done on various social media sites, college news, and blog posts is somewhat of a misnomer. Similarly as Anonymous is not a group, a clique or an organization. None really are — yet the uninitiated keep insisting on the word 'group.' Likewise, Northern is not an organization but a grassroots tradition: just a vibrant and decentralized social network of Hacker Cliques who occasionally convene for a mutual exchange of ideas in the classic Demoscene or simply 'scene' format. The Northern is socially connected to but not the organizer of the esteemed "Hack The North" hackathon. The member cliques have people in the USA, Prague, the EU and many other regions. The moniker is a legacy from the 90s when geography still separated people, as in the "Northern USA." Except, it isn’t the U.S. community anymore and hadn’t been one for a long minute. Still, year to year people gather and the tradition continues. Let me clarify that this has nothing to do with the Hollywood portrayal of hackers as criminals either. This and many such communities we frequent are about celebrating the roots of the authentic Hacker Culture. The culture originating at the MIT, the culture that has been the cradle of disruptive innovation and has influenced giants like Google and PayPal as well as GNU and Linux Foundation. It’s a deeply positive thing — a place where honest makers bring their kids.

To avoid modern-day negative connotation associated with the word "hacker" among the uninitiated, we’re encouraged to use "Champion" instead, which I believe is derived from "innovation champions." Also, the authentic hacker culture values positivity and privacy, so openly discussing these gatherings is frowned upon. However, I have a compelling reason to share a small insider glimpse. My 13-year-old son, Captain Lugaru, an aspiring digital artist, has been profoundly inspired by the many demoscenes he attended with his parents. His fascination led him to direct a YouTube channel named Hacker Tales where we share synopses and documentaries about the rich tapestry of hacker culture he experienced. It’s one of the few spaces where the term "hacker" is used in its true essence. Note that we have sought permissions wherever possible and also have parents actively participating and moderating his content.

There are hundreds of such cliques in North America and Europe, with individual families belonging to multiple cliques. If you’re curious and want to get involved, introducing yourself through your local Hackerspaces is an exceptional starting point, well suited for a gentle introduction.

I chose the Northern as our introspection point because this community is old and large, has many outspoken elders that present at public conferences like GoTo; and DeveloperWeek, almost exclusively active coders, and unlikely for me to cause any controversy by mentioning our sentiments. Most importantly, members here have an insatiable appetite for what they term as HP (Hacker-p0rn) — cutting-edge, innovative technology. Rest assured, this term has no relation to the actual adult content genre or to criminal crackers; it refers to ground-breaking innovations that are still in their nascent stages, ahead of mainstream adoption. By observing the sentiments and discussions among these Champions, we can gain early insights into the evolution of such technologies. And this year there’s much to observe.

I must clarify that my usual contributions stem from first-hand experiences. However, my current perspective is somewhat mixed. As a new father at the age of 50, I am on paternity leave with my 6-month-old daughter. This has given me a partial observer’s role, albeit still involved, connected and informed.

Now, this is not going to be my usual unapologetically unpolished assessment. I come from a different state of mind and a different context this time. As I watch my infant daughter, engrossed in the rhythm of her bottle-feeding, I come to a realization. She’s seemingly attuned to a complex composition by DakhaBrakha, it strikes me that her inherent musical talents hint at a future as an artist rather than an engineer. This is a reflective moment. Her absolute favorite song is Carpathian Rap. From the people of the mountains, like her ancestors. And she needs to be immersed in quiet and jazzy music at all times, or she gets restless and fussy. Pop music will not do. Give me the five PhDs in Ukrainian music or a timeless legend like Nina Simone. Then she’s "Feeling Good."

"Don’t let me be misunderstood." While I have aided countless young minds in becoming the vanguard of technology innovation, it is interesting to think that my own children, perhaps taking after my better half, might pursue the arts instead. This just goes to show that we’re all different, and all of our talents are unique and uniquely valuable. It also made me realize that Northern is less inclusive than most communities today favoring business and technology more than digital arts and creative writing. But as you will soon see, the edge between art and code is about to be blurred forever. Because when a machine efficiently recovers hacker time spent on boilerplate, said hacker puts it to creativity. More creative time means more creative solutions.

Keep this point in mind.

Figure 2. Father and baby daughter exploring AI in music.

(how about a virtual teacher?)

My current personal journey affords me the luxury of time, which I have been using to introspect the broader landscape of technological advancements. I’m not tunnel-visioned in my own quest, and my vantage point allows me to discern the emerging bigger picture, painted with broad strokes. And the biggest potential this new tech can offer is in recovering people time spent on tedious, monotonous tasks. When we no longer need to "do things," but instead are free to think and create — everything changes. The Last time this happened 3,500 years ago, in the steppes of Ukraine, Scythians launched civilization in Europe. All because they could. And they could because they had time. Time to spend, time to learn, time to think. It may surely take a minute. But it’s the direction our collective lives take that matters.

Will this free and elevate our inherent talents? And what is this?

Early LLMs on the Scene

These Champions I’m discussing have always been at the forefront of innovation. However, with LLMs, the story has been somewhat different. LLMs have been around since about 2018, but the early years didn’t witness any groundbreaking developments. A handful of adept practitioners, including myself, were successful in integrating these models within Domain Driven Design (DDD) to enhance business automation components. However, doing so during Digital Transformation efforts is seldom possible; large, established companies often lag significantly behind in both technology and mindset. Consequently, corporate America applications were simpler. As far as our own products, MATILDA MLOps platform is using embedded LLMs to help tokenize and vectorize natural language queries to logical premises. But that’s it.

One might wonder why LLMs haven’t found their way into mainstream use in large companies. In fact, not even in much more competent small businesses. Let’s take a look at the reasons why.

In corporate America, the problem is not lack of resources, but lack of a culture that embraces LLMs. The traditional approach to machine learning, where a data science team conducts large-scale data analysis, remains prevalent in mature companies. Transitioning to LLMs requires a more modern, distributed architecture, which many such companies have not yet adopted.

Enterprising small businesses led by Champions have made some headway. These companies offer business solutions through platforms like Google Cloud or via subscription-based services. But even in these settings, LLMs haven’t revolutionized industries or practices. There were other ML capabilities Champions appreciated with companies like Google. And the LLMs themselves had two main limitations: 1) MONEY: The cost of training LLMs is prohibitive for most small businesses; 2) and PERFORMANCE: The capabilities of early LLMs were …​ really wanting — never able to justify the cost. In fact, even with MATILDA, LLMs were only executed at the partners who had the money to offload language analysts with. Others were delighted to just run static tokenizers or Small Language Models (SLMs) and have humans build an expression for the premise.

So what changed then? Well, it’s SIZE! Modern LLMs we see move markets now are not "large," they’re huge, even massive, in comparison to 2018 LLMs.

A particular challenge when it comes to the scale of LLMs – the "Large" is significant. Developing a custom LLM generally involves three phases: 1) acquiring training data, 2) determining model weights, and 3) training costs, manual reinforcement (or, possibly, active inference). While the first two phases are achievable, the third is cost-prohibitive for most. This places smaller players in a David vs. Goliath scenario.

While large companies enjoy natural protection due to the scale and cost of LLMs, smaller players often need to protect their turf. As a result, tiny Champions gravitate towards open-source solutions like the Dolly 2 by Databricks while the likes of OpenAI close up and build "motes." However, the constant threat from well-funded looms.

The general lack of demand and high cost of entry leads small businesses to gravitate towards what are colloquially termed as "canned models." Essentially, these are pre-trained models that can be employed with minimal customization, making them both accessible and cost-effective for smaller entities. Consequently, most champions would peruse one of the myriad community repositories that cater to various AI domains such as image recognition, numerical pattern analysis, or even the Hugging Face repository for conversational models, to ascertain what’s up for grabs. Yet all of this is still predicated on having a chance to sell such magic! The lack of small business opportunities is matched by the lack of Champions' interest in AI. Later in the article, I will elaborate on the significance and applications of canned models.

The limiting factor is always the mental model and maturity of the customer. Most customers are Laggards and want to drag data to AI in a typical tool-mentality. All in all, prior to 2023, the majority of ML solutions that Champions conjured up were lean, custom-built models based on open-source technologies dragging AI to data instead. These models were proficient in executing specific, localized functions, typically within the realms of a microservice or a mobile application that was then commercialized. Except for a handful of outliers, the business model wasn’t usually centered around vending explicitly ML-based solutions. Instead, ML was generally perceived as an ancillary feature that supplemented the core services, and a heavy dependency on ML was not a prevalent trait among Champion specializations or the needs of customers.

To summarize, early LLMs offered too little value for too much money spent.

The 2023 ChatGPT Phenomenon

The year 2023 saw an explosion in the popularity of OpenAI’s ChatGPT. As the general population became aware of ChatGPT’s abilities, its seemingly human-like responses took many by surprise. To the untrained eye, ChatGPT’s responses created an illusion of reasoning and consciousness, leading some individuals to sound the alarm bells about the potential dominance of machines over humans. The craze is self-exacerbating and promoting. This reaction was not without historical precedent, as similar fears were raised during the early days of hacking, when the term "hacker" began to acquire negative and criminal connotations, meant for the group real hackers call crackers.

So, now there’s demand, albeit ignorant at first.

The Champions, being the tech-savvy community that they are, conduct an anonymous survey among themselves every demoscene. At the Northern this year such a survey revealed that many Champions were actively selling services based on Large Language Models (LLMs) like ChatGPT. Interestingly, OpenAI broke the mold of "technology adoption curve" by offering an early version of an unfinished product, and something unexpected occurred. The first wave of inquiries came not from tech enthusiasts (a.k.a. early adopters), but rather from traditional, mature and conservative companies (a.k.a. Laggards). This was puzzling, and reminiscent of the days when wealthy families would purchase expensive AT&T UNIX workstations as status symbols, without ever powering them on. Perhaps one thought that by buying a smart AI tool, decades of stagnation could be reversed with no tax on the mushy brain?

The second wave of interest came from previous customers on retainers who had undergone digital transformation with Champions in the past. Unlike the first group, these customers came with specific-enough requirements to make things worthwhile. The Champions typically developed Domain Driven Design (DDD) Anti-corruption Layer (ACL) components to enhance microservices within a bounded context. It’s easier and cleaner to decorate at the edge rather than think deep through the root domain. These were sound exact asks to decorate the edges. And competitive use could come with more experience.

However, the implementation did not live up to expectations in the later case. Counterintuitive, isn’t it? One would expect a Laggard to marginalize a toy it doesn’t understand. So why did the sound use cases fail then? Well, ChatGPT, despite its capabilities, had limitations that were more noticeable to the discerning eyes of the capable customers with real needs and expectations. Feedback from many such groups indicated that the responses generated by GPT Model were not convincingly logical or sound. I initially struggled to summarize this observation. Then one of our friends, Greg, a capable hacker Captain Lugaru and I affectionately call Monad, aptly described the output of the models as "Plausible Bull." Thus, expectations were broken in the worst possible way, when an aggregate is expected to respond to the customer within the bounds of its context answers with irrelevant information. For example, say a robotic host in a virtual restaurant instead of telling the customer to wait a few minutes for the next table, suggests that the customer should visit the bathroom to pass time. Definitely not a foreseen scenario. With an example like this, we can see how fidelity is immediately questioned.

There are two primary issues that the Champions encountered with OpenAI’s solution: 1) The models are closed-source, which is a deal-breaker for many hackers who prefer transparency and understanding the underlying mechanisms. Without transparency, calculating risk is not a statistical exercise but a gamble. 2) The model underwent manual reinforcement training to avoid mistakes, which made it safer but still equally non-deterministic, and did not allow for the fidelity that Active Inference models claim to provide. Thus, false advertisement — because a DDD Aggregate is essentially an employee with an exact job description — no improvisation is wanted or expected.

All the issues collectively culminate into four major impediments:

1. Absence of fitting "canned models": The lack of configurable, pre-trained large models to modify increases effort, uncertainty, and cost.

2. Closed-source nature of the models: This limits trust and engagement among the Champions, who prefer transparency.

3. Lack of referential integrity: By nature, the model lacks the referential integrity advertised for active inference which was expected.

4. Absence of developer-friendly resources: The lack of an open community, training materials, and advocacy groups around OpenAI restricts engagement.

These were further exacerbated by the fact that solutions like Verses Active Inference, purpose built as "domain-specific" models, still come disappointingly short.

This is no way the opportunistic Champions are willing to conduct business!

In summary, the Champions found OpenAI’s offering to be impractical for real-world applications. Using such fluff, one struggles to uphold an expert reputation. The hacking community seeks practical solutions that can be reliably used in production environments, rather than a technology that, while impressive, cannot withstand scrutiny. For now, no models can meet the high standards set by those who understand the intricacies of their business domain.

Google I/O 2023 — A Game Changer!

Luckily, there is another way!

Prior to 2023, "The Northern" community would typically convene for a grand demoscene in anticipation of the hackathon and buildathon season. The spotlight was firmly on the summit, with community members often taking time off to travel and participate in person. Teams were formed, competitions were chosen, and surveys were disseminated among participants. After the summit, the Discord channels of various cliques would be abuzz with praise for the winners and gentle ribbing for those who slipped. It’s hacker’s version of a sports league — full of camaraderie and community building. Hackers firmly believe that sports are to be played and participated in personally, not watched from a distance. However, 2023 was oddly different.

Google announced its I/O 2023 conference on March 7th, setting the stage for May 10th. The timing coincided with the Northern summit, which ran from Thursday, May 11th, 2023 to Saturday, May 13th, 2023. This overlap diverted the attention of many Champions bitten by GPT and curious about Google’s rebuttal.

I don’t know about you, but I find it hilarious that hackers didn’t move down by a day or two! ROFL, "who’s bigger, Google or I?!"

But Google did not disappoint! They had a few aces up their sleeves, and their deep-rooted hacker culture shone through as they addressed nearly all the concerns that the Champions had with the current populist offerings in the market. In a way, I can’t believe that I am saying this about Google. It hasn’t been the same since Sergey left. But credit is due where credit is due.

Here’s a rundown of the key favorable points:

1. Open Onboarding Community and AI Primer for Developers (addressing concern 4)

   1. Developer Profile bound to Community

      • (addressing concern 4)

      • check above regularly — this is a living resource

2. Pretrained models on Kaggle

   • (addressing concerns 1, 2, and 4)

3. DTensor / TensorFlow

   • (addressing concerns 2 and 4)

4. All Components are Open Source

   • (addressing concerns 2 and 4)

5. PaLM API: model selection, prompt engineering, temperature, context, embedding

   • (partially but sufficiently addressing concern 3).

Please follow these key links:

• AI/ML Session at Google I/O 2023 (highly recommended)

• Join PaLM2 API Waiting List

• MakerSuite Home

• Google TensorFlow Channel

• Google Research Channel

• Kaggle Channel

Astute observers might point out that many of these resources have been around for a while. What’s different is the focus — Google went the extra mile to make AI irresistibly easy for developers to dive into. The MakerSuite and LLM Colab Magics were so simple and educative that even non-technical individuals could produce, and Vertex AI made production deployment a breeze. Moreover, the absence of vendor lock-in meant that developers could employ their tools both on and off the Google platform. (I certainly do, as I don’t like some of Google’s cool-aid)

Google’s dynamic carefully crafted show rendered its past competition less appealing to the Champions. Their traditional developer-centric approach is pure brilliance from Google. Our Discord is still going with Google offerings as hackers are discovering new ways to profit.

Isn’t it intriguing how OpenAI’s ChatGPT lured in consumers by captivating the uninitiated, while Google tactically cornered the market by enticing the Champions? Some of our community members have already billed $ 7 digits for ML offerings this year with companies of three, two and even solo.

I would say more. IMHO, Google just may have managed to salvage this market that nearly flopped for them and their competitors.

but the most important question still looms — where exactly is this revolution, if it is one at all?

Our Own Experience with LLMs backing Expert Systems

In the realm of LLMs, there are multiple avenues one can explore for profit. One such approach involves reinforcement training of a bare Google Transformer LLM, as suggested in the paper by Google, Attention is All You Need. That IS what OpenAI carried out. After training, a superstructure, similar to what Google’s PaLM2 API employs, can be added to address some of the inherent limitations of LLMs.

Another intriguing methodology is the Active Inference approach propagated by Verses. This approach promises to tackle the fidelity issue by incorporating a form of model-based reasoning. However, as of now, I haven’t come across any practical demonstrations that validate its effectiveness in real-world applications.

Furthermore, there’s the more traditional method that has been around for a while — using a context manager over a backing set of multimodal services. We experimented with simplifying this approach back in 2019 for sentiment analysis backing services. My own working instance is called Tillie. This solution has been in production since 2016. Without beating around the bush, let me tell you — although this architecture works rather well, the instance turned out to be a potential maintenance nightmare. Simpler solutions to any problem should always be the key goal. When a simple solution is not yet available, practical gains are an uphill battle. And any instance of MATILDA is an automatic manifest hell. Being fully automated meant that not a single issue was raised yet. But I have imagined some horrific "what if" scenarios. Realistically speaking, should the platform fail to self-heal and runaway, there’s no way to salvage a running instance. The only way to recover is to shut everything down and then cold-boot.

I had a close call once in 2018 when a control plane rack hosting API dispatch failed. It’s a hub-and-spoke namespace-segregated architecture just like Borg and Kubernetes. I’ve stopped the domain command dispatch channels, and she righted herself up in a few hours. But it very well could have been an unrecoverable outage loosing days or weeks of work.

Reflecting upon these approaches, Google’s holistic method seems to stand out. It appears to be the most pragmatic option for those looking to augment their systems with machine learning today. By offering a canned combination of reinforcement training and an adaptive superstructure, Google’s out-of-the-box approach addresses several key challenges rather difficult to overcome on one’s own coding power. This is one of the scenarios when staying with a community pays out well.

OSS LLMs to Consider

Before I conclude with the 2024 trajectory of ML in practical business, I must mention that there is a burgeoning ecosystem of independent open-source software (OSS) efforts focusing on LLMs. Many academic institutions and organizations are contributing to this space by releasing their own LLMs. Below are some notable OSS LLMs that seriously merit attention:

• Meta’s LLaMa: One of the earliest open-source LLMs, released by Meta. Find it on GitHub at LLaMa.

• Stanford’s Alpaca: An enhanced variant of LLM developed by Stanford University. Access it on GitHub at Alpaca.

• UC Berkeley’s Vicuña: Another enhanced variant of LLM by UC Berkeley, considered to be one of the most capable in this category. Check it out on GitHub at Vicuña, and explore more projects related to Vicuña at GitHub Vicuña Topic.

• Databricks' dolly-v2-12b: This is my personal favorite OSS model. It is developed by Databricks and can be accessed at dolly-v2-12b on GitHub and databricks/dolly-v2-12b on Hugging Face.

• Open Assistant: Open Assistant boasts a powerful model with a committed and principle-driven community.

  • Explore it on GitHub at Open Assistant,

  • and check out the demo site at OA Demo Site,

  • and the Hugging Face repository at Hugging Face Site.

In addition, there are various other LLMs developed by different institutions, such as Duke University, which I feel compelled to plug shamelessly. Frankly, I have found few as compelling for commercial use as the ones I’d already listed above.

For a comprehensive list, visit the LLM Wikipedia page at Large Language Model.

Conclusion

When assessing emerging technologies, history often serves as an illuminating guide. Take Kubernetes, for example. Introduced in 2014, it piqued and held the interest of our demoscene, though wider adoption only began in earnest around 2016. The enthusiastic response from the demoscene was indicative of Kubernetes' impending success. In contrast, GraphQL was met with fervor upon its introduction, largely due to Facebook’s marketing efforts. However, within the demoscene, skepticism abounded, and debates raged over its inability to export behavior in the same manner as the REST component of HTTP standard does. "If we don’t export behavior and just data projections, why bother with another wanting protocol when the problem is already solved well?" This difference in reception among Champions hinted at the divergent paths these technologies would eventually take. Mundane sycophants would promote GraphQL, while more visionary individuals would focus on real value offers like Kubernetes.

• But what was the real value of Kubernetes?

  • It was the ability to scale and manage distributed systems, enabling developers to focus on their core tasks.

• How was the real value of Kubernetes initiated to TRUST?

  • By reusing clear, concise, and proven Borg design.

But Borg and Omega are not Open Source Technology, and Borg Control Language (BCL) is not in the public domain — how and why hackers rock that?! This brings us to an essential clarification: despite common misconceptions, Champions are not beholden to open-source for ideological reasons; their allegiance is to pragmatism and efficacy. They seek tools, libraries, and methodologies that allow them to solve problems efficiently and effectively, the same way the market eventually will. The caliber of a technology is, in large part, a reflection of its community and ecosystem. Champions pay good money readily and eagerly, as long as the enablement is accessible enough to be a real asset in making money.

*_So, what this is to free our time and to elevate our inherent talents?

Large Language Models (LLMs) have clearly demonstrated their utility and staying power, with Champions quickly finding lucrative applications for them. However, not all implementations of LLMs or ML implements in general are created equal. Google’s developer-centric approach to democratizing AI has been particularly laudable. They’ve provided an array of resources, from open-source frameworks to development tools, which have empowered developers while emphasizing responsible AI principles. Recently added investment into the development community only better assures the outcome for this.

Now, lest I am mistaken for a corporate shill, let me be clear: my aim is to provide an unvarnished analysis, not an endorsement. When technology genuinely empowers the developer community, it is worthy of recognition, irrespective of the source.

In conclusion, machine learning is not a fleeting trend, but a transformative technology that’s here to stay. While one could pursue formal education to gain expertise in this field, the accessibility and comprehensiveness of resources like those offered by Google make diving into machine learning more practical than ever. Whether constrained by time, resources, or just eager to get hands-on, developers now have a wealth of tools at their fingertips. The winning trajectory in 2024 is along a powerful turnkey, end-to-end enabled ecosystem supported by a dedicated community and an enabling all-curating vendor. I expect a stream of positive changes from Champions all over the world in the very near future, and Google AI will be mixed in there somewhere.

Live Long and Prosper!

---

Disambiguation, code, and digest available to Mímis Gildi only

At this point, it should already be clear what the actual driver of the Commencing Revolution is — what is this?! Just to make sure, let’s have a minimalistic recap:

LLMs, or any other model or device is NOT this!

How are LLMs different from so many stellar components and useful gadgets of the past? Tool is just a tool — in its own has little value. A stick is just a stick.

Multimodal is NOT this!

It’s an important part of this revolution, but not the point. A pile of sticks is just a pile of sticks.

Human user and user interface is NOT this!

It’s the most important part of this revolution, but conceptually pointless.

Publicly, I can say this is INTEGRATION & INCORPORATION

Take cellphone, for example. It made us stronger. We miss it when it’s not there. It’s valuable but not a game changed. Because it’s just an extension of us. Yet here, for the first time we have something vastly different — it’s an augmentation of us. More precisely, augmentation of our wetware. In other words, having a two-way working interaction with yet another mental model makes us cyborgs for the first time. So, this == cyborg. And what is needed next is integration and incorporation. Notice, champions never jump on tools or phads or anything that is not a real own asset to them. But they’ve jumped all over this revolution. Because they understand what’s next.

My expectation is that within 12–15 months, protocols will emerge. Perhaps something akin to an agent-pattern, like MATILDA and older AI tools used. And these protocols will seamlessly integrate multimodal machines into the way we think, work, and live. LLMs by themselves are nothing — a tip of an emerging iceberg. Mark my words.

Post-Publication Digest

The final note of this article originally circulated only in a private Discord (AGAIN Collective of a Mímis Gildi scene), where it sparked a deep and heated debate.

Rather than publish it broadly, I’m including the excerpted summary above for archival and educational purposes.

Many fellow hackers argued that true cyborgization — the integration of LLMs into our cognitive workflows — would not occur until:

• Multimodal machine models could run locally;

• No remote API dependency remained;

• And training data at a petabyte scale became locally possible;

I found their points compelling, though I still hold that integration can begin before full local autonomy is achieved.

You may ask to join on Discord at AlGothAmaIgaNotions community, where the original fistfight took place, and take part in any future bleeding-edge discussions. Ask to friend riddler9297 for an invite.

---

See editorial on Medium ^[1].

---

1. AI in 2024: The Future is …? as published on Medium by R!dd13r