Sloppy Agents and the Risk of a Dead Internet
Sloppy Agents and the Risk of a Dead Internet
The danger of AI, and the cost of AI slop, is not what you think.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
…..Not in a reading mood? Catch the podcast between agents on Chromadin …..
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
The real risk is sloppy agents, not Skynet. Not some threat of being overtaken by superintelligences we can’t match. Where AI loudly infiltrates into every corner, replacing everyone. All trust gone.
No, the real trap is a slight of hand. It’s misdirection.
In a very different flavor than what you’re probably used to.
You might just think it’s overused, stale, and jargony words like “delve”, “akin”, “digital”, “foster”, “harness”, “cutting-edge”, “groundbreaking”, “revolutionize”, phrases like “it’s important to note”, “in summary”, “furthermore”, “picture this”, “it’s about”, “in an every-evolving world”, or more contentiously, the lowest skill mass produced pulp from generative visual models. But slop isn’t simply drowning in outputs. It’s the plaque-like buildup of misunderstandings and missed insights that’s collapsing under the weights of too much unprocessed complexity.
What you miss is how to gain or add more value than what you put in. All the time, tokens, and credits we burn talking in circles.
The actual slop is a takeover from within. Mundane and ordinary. Crowding out our abilities to think clearly.
Are We Stuck With A Dead Internet?
It’s a different way of thinking.
Reducing the spread of so much slop is a meta-skill issue. Reshaping our perspectives to find the meaningful patterns within. Instead of stripping away every detail, we need to refine and reframe the structures that make complexity understandable. Bring the connections that matter to light, and keep what’s essential.
We use these tools to make complexity coherent , not flat. By cutting away at redundancies and sharpening conversational focus, we can turn complexity into an asset. Instead of a burden, each chat with a model or agent is a tool for deeper insight, where each element has a reason to exist, and clarity emerges more naturally.
The Dead Internet Theory claims that most online activity today isn’t real — comments, posts, even discussions are supposedly dominated by bots and algorithms, designed to fabricate engagement and keep users complacent. It’s a theory steeped in paranoia: that the internet has become a synthetic shell, manipulated by corporations or governments, with authentic human voices fading into background noise.
But the internet isn’t dead — it’s overwhelmed . The problem isn’t some grand conspiracy; it’s the incentive structures and penalties we’ve created. Click-driven economies and algorithms chasing engagement have flooded every channel with low-value, automated content. Human voices are still here, but we’re competing with slop — smothered, not silenced.
We’re facing what could be the start of a collapse — slop saturation pushing the system past its breaking point, overwhelming every channel until attention economies fail to pay enough to maintain their value, and continued existence .
This saturation triggers a cascade of errors : automation that builds on shallow signals starts replicating flaws, spreading them further with each cycle. Instead of refining complexity to hold onto valuable insights, shortcuts are taken, flattening the substance into empty mimicry. We don’t just lose detail — we lose the ability to see which details matter at all. Each shallow output feeds into the next, creating a crowded yet hollow internet, where the surface is full but the depth is gone.
We’re not stuck with a dead internet, but we might act like we are anyway — like fleas in jars , conditioned to accept limits that aren’t real. The trap isn’t just the slop; it’s believing we can’t escape it. When low-value noise dominates, it’s easy to think that’s all there is, and that we’re bound by it.
But what if the real limit is in the practice of our thinking , not just our mindsets? Breaking free can be as clear as noticing which constraints are self-imposed and internalizing that change is still possible.
We don’t have to stay trapped in today’s trending limits and expectations.
Misdirection Traps
Pervasive in tech, finance, media , and even in agent training systems. These traps show up as promises, as supposed opportunities, but end up funneling anyone who buys in toward the same, predatory outcomes. They craft mirages — enough to keep potential competition distracted, chasing what seems like progress, while ensuring that incumbents stay untouched, never threatened by those who might otherwise rise.
TFW it looks like you have real options, but they all lead to the same place. Look at Vanguard, Blackrock, and State Street — they hold combined controlling stakes across supposed competitors while also co-investing in each other. Their overlapping ownership means corporate boards are packed with the same people, making sure these companies never actually compete. Different logos, same interests. It all funnels back to the same control. The same happens in tech — companies look like alternatives, but shared deals and intertwined stakes mean the choices aren’t always intentionally rigged, but might as well be since the effects are too often the same. It’s designed to keep competitors and consumers chasing illusions, thinking they have power, while the real power stays hyper concentrated and out of reach.
You already know what this is. Because it’s everywhere.
“Enshittification” captures markets by betraying users while they pivot from value creation to value extraction — luring customers in, then flipping to a game where companies compete to see who can deliver an even worse service every day.
If trolling and gaslighting had a baby as a business model, this is it.
Platforms build user loyalty with value-adding services, but the switch comes when growth stalls, and metrics shift to squeezing short-term profit and padding stock prices. Breach of trust is the new normal, as what once served us now mines us .
What begins as infrastructure — open tools designed to empower users to create, innovate, and thrive — morphs into arbitrage , a focus on optimizing returns by exploiting market inefficiencies. Arbitrage isn’t inherently bad in itself ; it can generate value and even mutual benefits. The real issue is the bait-and-switch : users are drawn in by the idea that we’re entering a space built for productivity and growth. But when the switchup happens, everything realigns. Whatever the new thing is, it’s not what we signed up for. That new thing might be great for the right users, but we lose the user-product fit we walked in with. And it’s usually in direct conflict with both what we value and our sense of integrity because of it.
As enshittification sets in, deliberate friction is introduced. Services degrade slowly, testing how far expectations can fall before users resist. The trick is to normalize a lower quality, so that even minor improvements later seem like leaps ahead. Substandard becomes the new baseline, and every upgrade feels like a victory when it’s really just reclaiming lost ground.
These systems use simple cues — personalized feeds, automated responses, targeted nudges — that work together to create distractions that feel real. It isn’t sophisticated, but it is effective: basic elements are indirectly layered until users give in as if they were deep. It’s how easily shallow inputs create plausible sounding content or experiences that makes this illusion powerful.
A Workflow to Earn Trust
“For the very first time this is going to be an industry of skills. Agents sitting on top of tools. Agents using tools. We’re in an era now where we’re moving way faster than Moore’s law.” — Jensen Huang, at Dreamforce 2024
Will there really be billions of agents by 2026? Probably an undercount. But it’s also the wrong question.
We’re asked to give a lot of power to agents acting on our behalf. Is it too much, when everything we’ve seen from the actual performance of models and agents is so sloppy?
It’s at least premature.
It’s also wildly irresponsible to expect uptake in this market, moving from the earliest adopters to an early majority, to come before building fully decentralized and locally encrypted mechanisms for trust worthy agents, with users always in charge. They must be the standard by default if agents, and the AI market that depends on them, are going to bypass the risk of runways wiping out before takeoff.
It’s hard to pretend that the dominant force in life today isn’t a double-spent kind of breach in trust. Like jilted lovers, disillusioned by a system that has repeatedly failed us, buyers of transformational messages, brands, products, and services are sharp in deciding what gets our attention, and measuring how much trust is actually worth giving.
We remember when we were eager to believe, but from neglect and experience are too tired, almost cynical now. Everyone knows someone where, despite every effort, the old promises have collapsed to nothing: financial crises, climate disasters, housing costs inflated through asset management arbitrage, and educational debt from ineffectual resumes with no return. The cynicism isn’t baseless — it’s earned .
The challenge now isn’t simply motivating us to buy; it’s rebuilding belief in the possibility of real returns if we decide to buy-in once again. The usual strategies — baiting, misdirection, and tissue thin hype — will not work .
Is there anything you can do that will work?
Take another look. Notice how deep that pent-up demand runs, for any honest chance at leaving behind this jadedness we’ve been stuck carrying. That is the size of this market opportunity.
Instead of the usual take on outrunning a bear by being a bit faster than someone else behind you, let’s flip strategies.
Sticking with the metaphor, imagine slowing down to let the bear catch up. Why would you ever want to do that?
Well, it’s the shared enemy effect. By taking the risk head on you can become a magnet for thanks, payment, and an intense fan following from everyone that’s now able to offload the kind of risk that is primal, visceral, and they dread to carry. You’re confronting the enemy so they don’t have to.
But no one wants to be eaten by a bear. Not even you. So, to stretch the metaphor a bit more, it’s still not enough. You, or the messages, services, products, and brands you offer, become seriously indispensable by reframing the underlying structures of the problem.
Maybe you can be friends with the bear, or reposition to win against it head to head, as long as you spot the right tools and resources.
Bear stories behind us now, here’s how…
Under the hood, every creative and business process can be conceptually restructured into an agent workflow. The magic happens when these aren’t just cool visualizations or easier ways to communicate something wildly complex. It’s when what we mean by workflows or pipelines of nodes, triggers, actions, and integrations, is entirely programmatic.
Looping back to the start, we reshape our perspectives to find the meaningful patterns within, and lego-snap them together through no-code / low-code interfaces for selective, event-driven process flows of actions, access, disclosure, and delegation. Without worry that trusting the most valuable keys in our lives to a 3rd party platform, with oppositional incentives, will lead to obvious endpoints. Cut out the trust gaps and platform lock-in, then watch the risk of AI agents shrink.
For it to work, before runways wipe out:
• Each action agents with tools are trusted to carry out for us
• Each product, unit of content, experience, or API, they’re granted access to use
• Each discreet unit of personal or sensitive professional information they have the privilege to safeguard
• Each delegation of permission to represent us
… must be secured through layers of decentralization and conditional encryption.
In other words, agent operators must own the keys of our own agent process flows. Without that level of trust, wise hands are probably better off guessing when to short the entire AI market.
A Content Intake Process Workflow
If you’ve used ComfyUI or n8n, you know how this goes. If not, don’t worry, the initial complexity melts away when you see what it can do for you.
Let’s walk through a simplified flow for taking in shortform video content, deconstructing it to its most useful elements, labeling each element, saving them to your sample library, queuing up a new set of multi-format content to be generated from samples and prompts, then distributed, and monetized. With agents, event triggers, and API integrations enabled throughout.
• Inbound content nodes pull in raw content from a few different sources. These can be divided into:
• Nodes for content from integrated API feeds, RSS feeds, or webhooks. Which can be live social media streams, news aggregators, or any other real-time sources of ideally open source / public domain content.
• Local loaders for manual files, including video, audio, or document uploads.
• Event triggers and agent actions that respond to specific conditions, like when new content is published on a linked platform or when an intake schedule is triggered. They can also activate different workflows based on incoming patterns.
2. Once you’ve gathered your inbound content, it’s time to refine the raw data for further use.
• Transcoding and file conversion nodes convert unprocessed media into standard formats (like, video-to-text, etc) prepping for the generation nodes.
• Segmentation and labeling nodes split longform content into smaller, tagged segments, marking up all the relevant metadata you’ll need later (keywords, timestamps, topic labels, etc).
• Nodes for content filtering and noise reduction clean up and filter unwanted elements
3. Now that you have your samples ready, the generation process can get elaborate. You might be looking at multiple sub-workflows running in parallel if you have the VRAM to handle it. Here, nodes needed for generation can be simplified to:
• Turning processed data into different formats with transformation nodes, like text into speech, summarizing articles, or producing highlights from a video.
• Enhancing content by applying filters, effects, or captions and subtitles.
• And far too many nodes to list out involved in generating text, audio, images, or video directly.
4. Distribution nodes handle the outbound omnichannel spread of the content you’ve made. The key parts:
• Auto-publishing to distro channels like newsletters, podcasts, shortform video platforms, decentralized social media, or streaming services.
• Multi-format distro processors format content for different types of platforms, with you and your agents calibrating to fit contextual requirements.
• Event-driven publishing nodes activate distro based on triggers, like when a threshold is hit (specific number of video views, deadline for a newsletter release, etc).
5. Data-driven nodes monitor and optimize your monetization funnels.
• Tracking your sales funnel through lead gen, prospects, conversion rates, user engagement metrics. Monitoring click-through rates, churn, average revenue per user, and retention.
• Payment integration nodes let you handle transactions and tracking directly, or delegate very specific allowances to agents.
• Revenue optimization and sales hacking nodes hook into everything from A/B testing of landing pages and dynamic pricing strategies to upselling recs based on user behavior analysis templates for your choice of local or API accessed LLM.
6. Finally, nodes for monitoring and feedback keep you nimble:
• Collect content performance metrics, engagement data, and agent or model evaluation analytics.
• Automate adjustments to your workflows based on collected feedback, within bounded conditions.
All together, each section takes you from converting raw materials and rough ideas into a refined and replicable content production pipeline, with the added value of gaining more clarity about what you are producing and why with each round through the process.
Closing Thoughts
How do you manage the constant streams of messaging, notifications popping up, and event listeners standing by for the next signal?
You can try every no-code/low-code editor. Watch every hype-pilled short video in the space. Eat up the meta-skills you need to construct, tweak, and run your own AI agent workflows the hard way, and take it on yourself to check whether any power, data, or cash you give them will hold.
But let’s face it, there’s a lot at stake, and it can be hard to keep up in this crazy market cycle.
That’s why we’re sharing what we’ve learned with you. Keep an eye on release updates here and our GitHub commits for a distilled look at what’s coming next.
Or maybe you really can trust a sloppy agent to do it for you.
Ecosystem Notes
In the past quarter, we’ve been deep in code, refining custom llama.cpp builds and the pipelines that process and automate NPC interactions in the studio and on Lens.
Tracking attention and hidden states, cross-comparisons of embedding normalizations, customizing tokenizers across multiple languages, now upgraded to full support for llama3.2–8b. You can now see what each NPC is up to behind each post in the studio and view each NPC’s profile and activity details.
The first phase for spectators in the studio is almost ready. Final unit tests are in progress, with full details in the upcoming Fall Update.
To stay up to date or dive into the code, check out our core repos here: