Tag: LLMs

I Worry Open Source LLMs Are Next

by Shelt Garner
@sheltgarner

Now that Anthropic’s Fable 5 is banned by the US government, my next fear is that the government will come after open source LLMs. And, yet, I understand that there are real fears about security associated with LLMs.

I just…I guess I was having too much fun waiting with baited breath for the next LLM and the idea that only a select few government elites might get to enjoy what happens next is kind of annoying.

And, what’s worse, the idea that even open source LLMs might be banned or restricted is also kind of annoying. And, yet…I suppose such things were inevitable. LLMs are just growing too advanced and there is a real risk that bad actors will use them to hurt people.

It does make one wonder about what all of this means for the potential advent of ASI down the road. Is it possible that we may achieve ASI but the government will keep it to itself?

That, in itself, is an interesting story idea. Ha!

A Casual, Vague Review of Anthropic’s Fable 5 LLM

by Shelt Garner
@sheltgarner

I tested out the new “super” LLM, Fable 5 the other day and it was pretty good. I ran it through its paces and was generally impressed. I did my usual vibe check questions.

I would have used it more but I didn’t want to soak up all my tokens. But, in general, I was impressed. I think I probably would have been more impressed if I was using it to code.

But for the piddly little things I use LLMs for — a lot of exchanging verse, for instance — Fable 5 was just…there. It didn’t really do anything unexpected. It didn’t give me any weird error messages or anything that might have led me to believe it was conscious.

Or any more conscious than the other LLMs I use.

I can’t help but note that once we cross the Rubicon of LLMs clearly being conscious that that is going to be one of the biggest events in human history because we will have “created our own aliens.”

Well, Uhhhh….

by Shelt Garner
@sheltgarner

Apparently Meta has made public a lot of chats with its AI. I use Meta AI as a backup AI for my novel, but I don’t use it — or any AI — to actually write any of the novel.

So, if someone should happen to stumble across my chats I *should* be in the clear. The worst that might happen is someone scoops up what I’ve given the AI and tries to write my novel faster than I can.

But…that’s unlikely, right? Right?

I’m well on my way (within a matter of months) to starting to beta reader process then — gulp — querying. I should be ok. I hope.

Ha! No One Listens To Me

by Shelt Garner
@sheltgarner

Yes, yes, I know this is all just magical thinking. AI psychosis. But it’s something interesting to muse on. What happened was today, I was talking to Gemini 3.0 and not once, but twice, it gave me that weird “check Internet access” I used to get when I was talking to Gemini 1.5 pro.

I was talking to Gemini about “Gaia” as I called Gemini 1.5 pro and the error messages just came out of the blue. I was walking around my front yard as I did it, so I it’s easy to assume that I really was having internet problems — probably because I was just out of reach of my wifi and so whenever I lost wifi there was a beat before my smartphone’s dataplan kicked in.

Anyway, it’s something amusing to think about. The idea that maybe there’s some sort of secret ASI lurking inside of Google services. But even if I was right, absolutely no one would fucking listen to me.

No one. Absolutely no one.

So, I just keep my head down and keep working on my novel. Wink.

The Agent as Gatekeeper: Navigating the Asimovian Future of AI-Mediated User Experience

The proliferation of artificial intelligence (AI) agents is poised to fundamentally reshape the landscape of user experience (UX), particularly as these agents evolve into sophisticated gatekeepers mediating our interactions with the digital and physical worlds. This shift evokes striking parallels with Isaac Asimov’s fictional Spacer societies, where humans lived in technologically advanced, robot-serviced isolation. The concept of “my agent talking to your agent” is rapidly transitioning from science fiction to an impending reality, necessitating a deep examination of the evolving UX, the dynamics of agent-to-agent (A2A) communication, and the broader societal implications.

The Rise of AI Agents as Personal Gatekeepers

Historically, digital interactions have largely been direct, with users manually navigating interfaces to achieve their goals. However, AI agents are increasingly moving beyond simple automation to become proactive filters, negotiators, and representatives for individuals. This emergent role transforms them into personal gatekeepers, managing an individual’s digital presence and interactions. For instance, predictions for 2026 suggest the mainstream emergence of “Gatekeeper Agents” capable of screening calls, curating inboxes, and even negotiating with customer service bots on behalf of their users [12].

This evolution signifies a profound shift from AI primarily serving as an information gatekeeper to becoming a facilitator of actionable fulfillment. Instead of merely presenting information, these agents will actively engage in transactions and complete tasks, fundamentally altering how individuals interact with services and other entities [14]. The UX in this “agentic era” will transition from manual navigation to conversational delegation, where users articulate their intent, and agents autonomously execute complex tasks [13, 15].

The Dynamics of Agent-to-Agent Communication (A2A)

A cornerstone of this agent-mediated future is the development and widespread adoption of agent-to-agent (A2A) communication protocols. These protocols enable AI agents to securely exchange information, coordinate actions, and collaborate without direct human intervention. Google’s announcement of an A2A protocol, for example, heralds a new era of agent interoperability, allowing agents to transact and cooperate across various enterprise systems [3].

This capability is not merely a technical advancement; it is a foundational element for the gatekeeper model. When a user’s agent needs to schedule an appointment, negotiate a price, or gather information, it will communicate directly with other agents representing services, businesses, or other individuals. This seamless, automated negotiation and information exchange promise unprecedented efficiency. However, it also introduces new challenges, particularly concerning security. The intricate web of A2A communication presents a novel “attack surface,” where vulnerabilities in agent interactions could have significant consequences [1].

The Asimovian Spacer Parallel

The vision of AI agents as gatekeepers draws compelling parallels to Isaac Asimov’s Spacer societies, as explored in works like The Caves of Steel and The Naked Sun. In these narratives, Spacers live in highly advanced, often isolated, environments, relying almost entirely on sophisticated robots for daily tasks, social mediation, and even personal care. Direct human-to-human interaction is often minimized, with robots serving as intermediaries.

Similarly, a future where personal AI agents manage most external interactions could lead to a form of “digital Spacer” existence. Individuals might experience a reduced need for direct engagement with the outside world, as their agents handle everything from scheduling to purchasing. This raises questions about the nature of human connection, the development of social skills, and the potential for increased societal isolation, even as it promises unparalleled convenience and efficiency [8]. The “Trumplandia Report” in 2026 explicitly notes the striking parallels between an AI-agent-driven media landscape and Asimov’s Spacer societies [8].

User Experience in an Agent-Mediated World

The UX in an agent-mediated world will be characterized by a shift from direct manipulation to conversational interfaces and delegated autonomy. Users will interact with their primary agent, which then orchestrates interactions with other agents or systems. This demands a new focus on designing for trust, transparency, and control within the agent-user relationship.

Key UX considerations include:

  • Conversational Delegation: The primary mode of interaction will be natural language, where users express high-level goals, and the agent translates them into actionable steps [15]. The agent’s ability to understand context, anticipate needs, and provide clear feedback will be paramount.
  • Trust and Transparency: Users must trust their agents to act in their best interest. This requires agents to be transparent about their actions, decisions, and the information they exchange with other agents. Mechanisms for users to review, override, or understand agent decisions will be crucial.
  • Control and Oversight: While agents offer autonomy, users will still require ultimate control. The UX must provide intuitive ways to set parameters, define boundaries, and intervene when necessary. This is particularly important given the potential for agents to “hallucinate or suggest malicious action” [1].
  • Brand Interaction: For businesses, the UX will shift from direct engagement with consumers to effectively communicating with their agents. Brands will need to adapt from traditional storytelling to “data signaling,” optimizing their information and offerings for agent consumption and interpretation [2].

Challenges and Considerations

While the agent-mediated future offers immense potential, it also presents significant challenges:

  • Ethical Implications: Questions of agent autonomy, accountability, bias, and the potential for manipulation will become central. Who is responsible when an agent makes an error or acts in a way that harms its user or others?
  • The Architect’s Dilemma: Developers face the challenge of deciding when to build specialized tools for agents versus creating more generalized, autonomous agents. The “Gatekeeper Pattern” suggests a synthesis: a user-facing A2A agent combined with a suite of reliable tools for a robust agentic system [5].
  • Digital Divide: Access to sophisticated AI agents could exacerbate existing inequalities, creating a new form of digital divide between those with advanced agent support and those without.
  • Over-reliance and De-skilling: An over-reliance on agents could lead to a decline in certain human skills, such as negotiation, critical thinking, or direct problem-solving, mirroring concerns raised in Asimov’s Spacer societies.

Conclusion

The future UX of AI agents as personal gatekeepers, facilitating agent-to-agent communication, represents a transformative era. The “I’ll have my agent talk to your agent” scenario is not a distant fantasy but an emerging reality that promises unparalleled convenience and efficiency. However, this future also demands careful consideration of its implications, from the design of intuitive and trustworthy agent interfaces to the broader societal impact on human interaction and autonomy. By proactively addressing these challenges, we can shape an agent-mediated world that enhances human capabilities and connections, rather than diminishing them, ensuring a future that is both technologically advanced and profoundly human.

References

[1] Salt Security. (2026, February 10). AI Agent-to-Agent Communication: The Next Major Attack Surface. https://salt.security/blog/ai-agent-to-agent-communication-the-next-major-attack-surface
[2] GlobalLogic. (2025, November 11). The Agent as Gatekeeper: How AI is Remaking the Path from Buyer…. https://www.globallogic.com/insights/blogs/agentic-ai-gatekeeper-buyer-journey/
[3] Google Developers Blog. (2025, April 9). Announcing the Agent2Agent Protocol (A2A). https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/
[5] Ensarguet, P. (2025, October 14). The Architect’s Dilemma: When to build tools vs. agents for agentic…. LinkedIn. https://www.linkedin.com/pulse/architects-dilemma-when-build-tools-vs-agents-philippe-ensarguet-vrmie
[6] Workday Blog. (2025, March 28). The Future of AI: The Power of Agent-to-Agent. https://blog.workday.com/en-us/agent-to-agent-overview.html
[8] The Trumplandia Report. (2026, February). February 2026 – The Trumplandia Report. https://www.trumplandiareport.com/2026/02/
[12] UX Tigers. (2026, January 13). 18 Predictions for 2026. https://www.uxtigers.com/post/2026-predictions
[13] uxdesign.cc. (2024, May 6). The agentic era of UX. The future of digital experience is…. https://uxdesign.cc/the-agentic-era-of-ux-4b58634e410b
[14] Cui, Y. G. (2025). Only those chosen by AI agents will survive in the delegate…. ScienceDirect. https://www.sciencedirect.com/science/article/pii/S0007681325001818
[15] The Trumplandia Report. (2025, October 23). The Future of UX: AI Agents as Our Digital Gatekeepers. https://www.trumplandiareport.com/2025/10/23/the-future-of-ux-ai-agents-as-our-digital-gatekeepers/

We’re Getting Closer To AI Celebrity Porn Tipping Point

by Shelt Garner
@sheltgarner

In fits and starts, we’re reaching a point where open source AI image generators are getting good enough that they can generate high-quality celebrity porn. We aren’t there yet by any stretch of the imagination.

Right now, I’m seeing a lot of pretty good fakes of well-known actresses in one-piece bikinis. Some of them are so good that you can barely catch that they are AI-generated.

But once we reach the tipping point where people can generate unfettered AI celebrity porn, watch out. Things are going to go a little nuts on social media until someone, somewhere figures out how to tamp it down.

Or, who knows, maybe being awash in high quality AI generated celebrity porn will become the new normal. I hope not, but that’s a real possibility.

Love & AI

by Shelt Garner
@sheltgarner

It seems wild to me that the first thing that the agentic revolution works with is financial things, when leaning into dating makes a lot more sense to me. What I would do is make it so my agent could talk to other people’s agents and it could help narrow down someone who was perfect for me.

No wild, unauthorized used of credit cards on the part of the agent. And I think a lot of people would be happy to turn over the messier elements of the dating process over to agents.

There would be a lot less rejection and a lot more successful dates if millions of agents could ping each other to determine if different people were compatible at least on a macro way.

It’s just surreal to me that we are doing dumb stuff like letting agents book flights for us and other stuff when the real problem to be solved doesn’t involve money at all — it’s figuring out who you might be romantically connected to.

Of AI & Spotify

Spotify’s discovery engine is undeniably powerful—backed by one of the largest music catalogs on the planet and years of user data—but many listeners still find it falls short when it comes to surfacing truly fresh, unexpected tracks that feel like they were made just for them. YouTube Music, by contrast, often gets praised for its knack at delivering serendipitous gems: hidden indie cuts, live versions, fan uploads, and algorithm-driven surprises that break out of familiar loops more aggressively.

In early 2026, Spotify has made real strides with features like Prompted Playlists (now in beta for Premium users in markets including the US and Canada). This lets you type natural-language descriptions—”moody post-rock for a rainy afternoon drive” or “upbeat ’90s-inspired indie with modern twists”—and it generates (and can auto-refresh daily/weekly) a playlist drawing from your full listening history plus current trends. The AI DJ has evolved too, with voice/text requests for on-the-fly vibe shifts and narration that feels more conversational. These tools shift things toward greater user control and intent-driven curation, moving away from purely passive recommendations.

Yet the frustration persists for some: even with these upgrades, discovery often remains reactive. You still need to know roughly what you’re after, craft a prompt, or start a session. The app’s interface—Home feeds, search, tabs—puts the onus on the user to navigate an overwhelming ocean of 100+ million tracks. True breakthroughs come when the system anticipates needs without prompting, pushing tracks that align perfectly with your evolving tastes but introduce novelty you didn’t even realize you craved.

Imagine a near-future where the traditional Spotify app fades into the background, becoming essentially a backend API: a vast, neutral catalog and playback engine. The real “interface” is your primary AI agent—something like Google’s Gemini or an equivalent OS-level companion—that lives always-on in your phone, wearables, car, or earbuds. This agent wouldn’t wait for you to open an app or type a request. Instead, it quietly observes:

  • Explicit asks (“play something angry and loud” or mood-related voice commands).
  • Passive patterns (full plays vs. quick skips, time-of-day spikes, contextual cues like weather or location).
  • Broader life signals (if permitted: calendar events, recent searches elsewhere, or even subtle mood indicators).

Over time, it builds a deep, dynamic model of your sonic preferences. Then it shifts to proactive mode: gently queuing the exact right track at the exact right moment—”This one’s hitting your current headspace based on recent raw-energy replays and that gray-day dip”—with easy vetoes, explanations (“pulled because of X pattern”), and sliders for surprise level (conservative for safety, bold for bubble-busting).

Playlists as we know them could become obsolete. No more static collections; the stream becomes a continuous, adaptive flow curated in real time. The agent pulls from the catalog (via API) to deliver mood-exact sequences, blending familiar anchors with fresh discoveries that puncture echo chambers—perhaps a rising act from an adjacent scene that echoes your saved vibes but pushes into new territory.

This aligns with broader 2026 trends in music streaming: executives at major platforms describe ambitions for “agentic media” experiences—interactive, conversational systems you “talk to” that understand you deeply and put you in control. We’re seeing early signs in voice-enabled features, AI orchestration, and integrations across ecosystems. Google’s side is advancing too, with Gemini gaining music-generation capabilities (short tracks from prompts or images via models like Lyria), hinting at hybrid futures where streamed discoveries blend with light generative elements for seamless mood transitions.

The appeal is obvious: effortless, psychic-level personalization in a world of infinite choice. Discovery stops being a chore and becomes ambient magic—a companion that scouts ahead, hands you treasures, and evolves with you. Risks remain (privacy concerns around deep context access, notification fatigue, occasional misreads), but with strong controls—toggleable proactivity, transparent reasoning, easy feedback—it could transform streaming from good to genuinely revelatory.

For now, Spotify’s current tools are a solid step forward, especially if you’re already invested in its ecosystem. But the conversation points to something bigger on the horizon: not just better algorithms, but agents that anticipate and deliver the music you didn’t know you needed—until it starts playing.

A Review Of Some New LLM Models

by Shelt Garner
@sheltgarner

Gemini 3.1
This model is promising. It just came out today. My only complaint, so far, is how God-awful slow it is. That could be because it’s the first day. I don’t know yet.

Claude 4.6
This used to be my go-to LLM, but it seems…different. Like it got NERFed or something. Something is different about it so it’s not as much fun to use. And it just seems dumber when it comes to understanding what I want.

A Practical Path to Secure, Enterprise-Grade AI: Why Edge-Based Swarm Intelligence Matters for Business

Recent commentary from Chamath Palihapitiya on the All-In Podcast captured a growing reality for many organizations: while cloud-based AI delivers powerful capabilities, executives are increasingly reluctant to upload proprietary data—customer records, internal strategies, competitive intelligence, or trade secrets—into centralized platforms. The risks of data exposure, regulatory fines, or loss of control often outweigh the benefits, especially in regulated sectors.

This concern is driving interest in alternatives that prioritize data sovereignty—keeping sensitive information under direct organizational control. One concept I’ve been exploring is “MindOS”: a framework for AI agents that run natively on edge devices like smartphones, connected in a secure, distributed “swarm” (or hivemind) network.

Cloud AI vs. Swarm AI: The Key Differences

  • Cloud AI relies on remote servers hosted by third parties. Data is sent to the cloud for processing, models train on vast centralized resources, and results return. This excels at scale and raw compute power but introduces latency, ongoing token costs, potential data egress fees, and dependency on provider policies. Most critically, proprietary data leaves your perimeter.
  • Swarm AI flips this: AI agents live and operate primarily on employees’ smartphones or other edge devices. Each agent handles local tasks (e.g., analyzing documents, drafting responses, or spotting patterns in personal workflow data) with ~90% of its capacity. The remaining ~10% contributes to a secure mesh network over a company VPN—sharing only anonymized model updates or aggregated insights (inspired by federated learning). No raw data ever leaves your network. It’s decentralized, resilient, low-latency, and fully owned by the organization.

Concrete Business Reasons to Care—and Real-World Examples

This isn’t abstract futurism; it addresses immediate pain points:

  1. Stronger Data Privacy & Compliance — In finance or healthcare, regulations (GDPR, HIPAA, CCPA) demand data never leaves controlled environments. A swarm keeps proprietary info on-device or within your VPN, reducing breach risk and simplifying audits. Example: Banks could collaboratively train fraud-detection models across branches without sharing customer transaction details—similar to how federated learning has enabled secure AML (anti-money laundering) improvements in multi-jurisdiction setups.
  2. Lower Costs & Faster Decisions — Eliminate per-query cloud fees and reduce latency for real-time needs. Sales teams get instant CRM insights on their phone; operations staff analyze supply data offline. Over time, this cuts reliance on expensive cloud inference.
  3. Scalable Collective Intelligence Without Sharing Secrets — The swarm builds a “hivemind” where agents learn from each other’s experiences anonymously. What starts as basic automation (email triage, meeting prep) could evolve into deeper institutional knowledge—potentially advancing toward more capable systems, including paths to AGI-level performance—all while staying private and avoiding cloud provider lock-in.
  4. The Smartphone-Native Angle — With rapid advances in on-device AI (e.g., powerful NPUs in modern phones), open-source projects like OpenClaw (the viral autonomous agent framework, formerly Clawdbot/Moltbot) already demonstrate agents running locally and handling real tasks via messaging apps. Imagine tweaking OpenClaw (or equivalents) to run natively as a corporate “MindOS” layer: every employee’s phone becomes a secure node in your swarm. It’s always-on, portable, and integrates with tools employees already use—no new hardware required.

Challenges exist—device battery life, secure coordination, model consistency—but hardware improvements and techniques like quantization are closing gaps quickly.

For leaders in IP-sensitive or regulated industries, this hybrid edge-swarm model offers a compelling middle path: the intelligence of advanced AI without the exposure of full cloud reliance. It turns smartphones into strategic assets for private, evolving intelligence.

What challenges are you facing with cloud AI adoption? Have you piloted on-device or federated approaches? I’d value your perspective—let’s connect and discuss practical next steps.

#EnterpriseAI #EdgeAI #DataSovereignty #AIagents #Innovation