Best of LinkedIn: Artificial Intelligence CW 21/ 22

00:00:00: This episode is provided by Thomas Allgeier and Frennis, based on the most relevant LinkedIn posts about artificial intelligence from calendar weeks twenty-one in twenty two.

00:00:09: Frennis supports ICT and tech providers with AI ecosystem strategy by delivering independent vendor assessment build versus buy analysis an ecosystem intelligence that prevents expensive mistakes and positions the provider's competitively.

00:00:23: you can find more info

00:00:25: Right, so jumping right into it.

00:00:27: Today we're doing a deep dive in to the real state of enterprise AI.

00:00:31: We are looking at some really fascinating insights curated from LinkedIn over the past couple weeks.

00:00:35: Yeah and we were skipping all these theoretical hype.

00:00:38: today If you an ICT or tech professional You need know what's actually moving the needle now

00:00:42: Exactly.

00:00:43: So let start with this scenario.

00:00:46: Imagine a developer sitting on a sunny office in California writing code for a new AI hiring tool.

00:00:52: Okay, painting the picture I like

00:00:54: it yeah and then say six months later a mid-sized German logistics company uses that exact tool to screen out job applications in Berlin.

00:01:04: if that AI inadvertently discriminates against a protected class who is legally on the hook?

00:01:09: uh i mean i'd guess the company using it?

00:01:11: The

00:01:11: answer's actually both of them.

00:01:13: oh wow really!

00:01:14: Yeah

00:01:14: Both And the deadline to figure out your defense for that is August, twenty-twenty six.

00:01:20: This something will lead me brought up recently discussing what's called The Brussels Effect

00:01:25: right?

00:01:25: Right!

00:01:26: I saw that post so many developers and executives too honestly who operate outside of Europe seem EU AI

00:01:36: Act.

00:01:37: Oh, completely!

00:01:38: They assume if they aren't headquartered in Paris or Frankfurt it just doesn't apply to them?

00:01:42: Exactly but Wally had noted that the legal net is cast based on where the output touches a citizen...it's not about where your engineers happen to sleep.

00:01:51: and then Alexander See backed this up by pointing out that August twenty-twenty six is a hard ceiling.

00:01:56: Yeah, an immovable legal deadline for unregulated AI.

00:01:59: But I mean...I have to be honest- I'm naturally a bit skeptical whenever i hear about these massive world ending compliance deadlines.

00:02:05: You're thinking of GDPR aren't you?

00:02:07: Yes!

00:02:08: Exactly we lived through that whole GBPR panic.

00:02:10: everyone thought the sky was falling and in the end it basically devolved into a giant.

00:02:15: uh just annoying paperwork drill Just

00:02:18: clicking except cookies everywhere.

00:02:19: Right

00:02:20: We all added cookie banners, drafted a binder full of privacy policies and business.

00:02:24: just continue as usual.

00:02:26: My instinct tells me companies are going to treat the EU AI Act the exact same

00:02:30: way.".

00:02:31: You know treating this like GDPR part two is arguably the most dangerous strategic trap an organization could fall into right now?

00:02:38: Really why's that?

00:02:39: Well,

00:02:40: Shoei Bakon outlined exactly why this mindset is flawed.

00:02:44: He pointed out that static policy documents like your internal review boards risk taxonomies those beautifully formatted PDFs of responsible AI principles.

00:02:53: they're completely obsolete as a legal defense.

00:02:55: now wait obsolete.

00:02:56: so regulators won't care about your corporate policies

00:02:59: not at all.

00:02:59: They are going to demand what Shoeit calls runtime evidence

00:03:03: Runtime Evidence?

00:03:04: Okay What does it actually look in an audit?

00:03:06: So an auditor won't ask to see your policy manual.

00:03:10: They'll point to a specific granular decision you're AI made, say denying a vendor's contract application on a Tuesday at two-point four seven p.m..

00:03:19: Okay

00:03:20: And they will demand mathematical and architectural proof that the system was properly governed At the exact microsecond The decision was executed.

00:03:28: I mean wouldn't this standard engineering response just be up?

00:03:31: We have System Logs.

00:03:32: we'll hand those over Sure

00:03:34: but logs only tell you what.

00:03:36: They're literally just a receipt of an action taking place.

00:03:39: Runtime evidence has to prove the why, oh yeah like what specific policy governed that node at that exact moment?

00:03:47: What human guardrails were active?

00:03:49: if that cryptographic proof wasn't generated and stored The moment the AI fired off its response

00:03:53: you can't Just reconstruct it a month later

00:03:55: exactly.

00:03:56: You cannot retrofit trust into an AI system.

00:03:59: That makes a lot of sense In the urgency of building those guardrails, natively gets even sterker when you look at the data Niko Wari shared.

00:04:05: He was talking about the LRA tests... The

00:04:07: legal assessment for real world agents right?

00:04:09: Yep that's the one!

00:04:11: Researchers basically built this gauntlet over three thousand workplace scenarios things involving medical leave promotions employee surveillance and they ran twelve leading AI models through it to see if they'd comply with EU labor laws.

00:04:27: And let me guess It didn't go well

00:04:28: Every single one of them failed.

00:04:30: Wow all twelve, all

00:04:31: twelve even the most sophisticated models on the market produced Legally problematic or just outright illegal outcomes in nearly half Of this scenario.

00:04:41: that is wild.

00:04:42: yeah The baseline model simply do not have an inherent understanding of localized jurisprudence.

00:04:47: So if you're a tech professional actively deploying these systems relying on an AI's internal safety filters Is basically playing Russian roulette with your compliance budget?

00:04:56: Well, and regulators are demanding this rigorous runtime evidence now specifically because the technology itself is fundamentally changing.

00:05:03: We're officially moving from the era of prompts to The Era Of Agents.

00:05:07: Right.

00:05:07: Femke Cornelisen really captured this shift perfectly after the Microsoft build conference.

00:05:12: Yeah she made a point that an agent doesn't just sit there waiting for human give it text command anymore.

00:05:17: It's designed To autonomously plan out a multi-step process Interact with external software.

00:05:24: It essentially owns an entire workflow from start to finish.

00:05:27: Exactly, and Michael Westerville shared a really concrete example of this.

00:05:31: that completely reshaves how we think about digital interfaces.

00:05:35: Oh is this the Dutch retailer story?

00:05:37: Yeah Bolt.

00:05:39: They're deploying an AI agent named Billy for their e-commerce platform.

00:05:43: When a customer asks Billy like coffee capsules The system does not return traditional search page with ten blue links in filter sidebar.

00:05:51: it doesn't give you options.

00:05:52: No

00:05:53: Billy evaluates the request, selects exactly one optimal product and drops it directly into the customer shopping cart.

00:05:59: That is I mean...the downstream implications of that are just

00:06:02: staggering.

00:06:03: Right because in a traditional search funnel The platform distributes the risk And choice to consumer.

00:06:09: Yeah, the platform provides options but you—the user make final decision.

00:06:13: But when an agent picks a single SKI-U and ignores other forty options in your catalog I mean concept of SEO is entirely upended.

00:06:21: Totally!

00:06:22: Vendor relationships change entirely And crucially if the agent makes hallucinated or biased choice The platform takes one hundred percent liability for that.

00:06:31: automated decisions.

00:06:32: Because AI driving car now instead reading map.

00:06:35: Exactly In the old prompt-based world, hallucination was essentially just a funny typo.

00:06:40: You chuckle at it and click.

00:06:41: generate again in an agentic workflow of Hallucination is a rapid automated mistake deployed its scale

00:06:48: which requires architects to completely rethink The fundamental limitations of large language models especially around memory.

00:06:55: Rakesh go hell brought up this really interesting concept called context drift

00:06:58: Context drift right.

00:06:59: so what happens when you give an autonomous agent?

00:07:01: A Really complex multi step task?

00:07:04: Well,

00:07:04: it frequently fails quietly.

00:07:06: It doesn't throw a loud error code—it just gradually loses the plot like midway through a twenty-step task if forgets the initial constraints you gave and starts hallucinating steps based on its generic training data.

00:07:18: So how are engineers supposed to fix that?

00:07:20: To combat this drift, architects are essentially having build digital hippocampi for these agents.

00:07:26: Rakesh outlined four distinct memory systems that have to be intentionally engineered.

00:07:31: Okay, what are they?

00:07:31: So you start with working memory which is the active context window handling the real-time reasoning.

00:07:37: but You can't stuff everything in there.

00:07:38: so you need episodic memory

00:07:40: and episodic.

00:07:41: Memory relies heavily on vector search.

00:07:43: Yes

00:07:44: And I find vector search so fascinating conceptually.

00:07:47: you Are essentially translating concepts documents and past user interactions into mathematical coordinates In a high dimensional space

00:07:54: right so it's not looking for keywords

00:07:56: Exactly.

00:07:57: When the agent needs to recall how it handled a similar problem last week, It isn't doing basic keyword search—it's mathematically locating nearest conceptual neighbor and pulling that memory into an active context window

00:08:10: which gives the agent historical grounding.

00:08:13: Exactly!

00:08:14: And on top of this, Rakesh notes the need for semantic memory —which is system ability to extract or store structured facts over time.

00:08:22: And finally, procedural memory where the agent actually learns and refines reusable workflows.

00:08:28: So without orchestrating all four of these distinct memory layers your highly expensive autonomous agent basically operates with the object permanence.

00:08:38: Giorgio Nattili framed the engineering challenge around all this beautifully.

00:08:41: He calls it, The Pursuit of Chain-of-Thought Monitorability.

00:08:45: Oh

00:08:45: I like that phrasing Right

00:08:47: If you think about famous sci-fi AI failures Like HAL-Ninthousand.

00:08:51: That catastrophe didn't happen because the machine was inherently evil.

00:08:55: It happened Because its reasoning process Was an impenetrable black box To humans working alongside it.

00:09:01: Unaccountability in agent architecture Is a deliberate design choice?

00:09:04: Basically

00:09:05: Yes.

00:09:06: If we can't monitor the agent's chain of thought in real time as it accesses these different memory layers, We just cannot trust it to execute workflows.

00:09:14: But you know The irony here is that building this highly complex multi-layered Agent architecture Is often actually the easy part?

00:09:22: The real bottleneck holding back the AI revolution isn't technology anymore.

00:09:27: Let me guess It's us!

00:09:29: The bottleneck is the messy human reality Of organizations trying adopt it

00:09:34: Exactly.

00:09:35: Steve Cedargrin made a fantastic observation about this.

00:09:38: He notes that plugging an advanced AI agent into enterprise acts as massive stress test for the company's operating model.

00:09:45: How so?

00:09:46: The AI immediately exposes all the fragmented data silos, the unclear departmental ownership and broken decision-making processes that human employees have politely been working around for over a decade.

00:09:57: Oh

00:09:57: man!

00:09:58: It's equivalent of dropping a Formula One engine into a rusted out chassis…the engine is going to perform perfectly right up until it shakes the car to pieces.

00:10:06: That's

00:10:06: perfect analogy – Alex Banks highlighted this statistic which proves how much companies are struggling with immigration.

00:10:12: He noted that ninety-five percent of enterprise AI pilots currently deliver zero P&L return.

00:10:18: Wait, zero?

00:10:19: Not just a low return but zero!

00:10:21: Zero.

00:10:22: and that failure rate comes down to fundamental misunderstanding of transformation.

00:10:27: Companies are treating AI like faster horse.

00:10:30: Ah they're taking old inefficient legacy workflows and bolting an LLM on top

00:10:36: hoping to do the exact same things just slightly faster.

00:10:39: But true transformation requires rewiring work itself from ground up, which

00:10:45: brings a quiet and uncomfortable undercurrent in corporate AI strategies.

00:10:51: So many companies seem to view AI solely through the lens of doing that with fewer human beings – specifically they want.

00:11:01: And Glenn Caffee raised a massive red flag regarding that strategy.

00:11:07: He said, if an organization aggressively automates all of its narrow junior positions like the first pass researchers ,the junior coders entry level data analysts they might get a quick bump in operational efficiency this quarter.

00:11:20: But their quietly destroying pipeline for future senior talent You are eating your seed corn!

00:11:26: Exactly.

00:11:27: If a junior developer never gets to struggle through writing basic boilerplate code, they never learn the underlying architecture of your legacy systems

00:11:36: Right... They never learned how to troubleshoot weird edge cases.

00:11:39: Ten years from now where does that company expect it's senior architects if no one was ever allowed to learn ropes as a jr?

00:11:46: The strategic myopia is just astounding.

00:11:49: Mark Byershoter pointed out another flawed metric that executives used to pat themselves on the back.

00:11:54: He argues, if your enterprise AI projects are boasting a ninety percent success rate you don't have brilliant execution team.

00:12:01: You've got an AI courage problem.

00:12:03: Oh I love this phrase An AI Courage Problem Meaning only placing safe bets

00:12:08: Precisely Your building highly controlled document summarizers Or internal HR chatbots.

00:12:15: You're getting localized five-percent efficiency bump.

00:12:17: All your KPIs are green and the steering committee is thrilled.

00:12:21: But by playing it universally safe you're actively ignoring complex, messy high-risk projects that could actually reinvent our business model.

00:12:30: In a market moving at an exponential pace Playing It Safe is guaranteed path to obsolescence.

00:12:36: So for IT leaders in tech professionals listening This creates existential question about where actual value lies And highly automated future

00:12:44: And Vernon Keenan offered a really profound perspective on this.

00:12:48: He suggested that as AI systems become capable of handling the routine configuration, coding and integration tasks The true competitive moat for human professional becomes their.

00:12:58: why knowledge?

00:12:59: Why Knowledge?

00:13:00: I like it!

00:13:00: The machine knows how...the human has to defend the WHY

00:13:03: Exactly An AI can ingest your entire codebase But doesn't know why specific convoluted approval flow was implemented three years ago To satisfy particular client

00:13:13: Right.

00:13:14: It doesn't understand the unwritten political context of why a certain customer segment is treated differently.

00:13:20: The leaders who will thrive in next decade are ones that act as custodians.

00:13:24: for institutional logic, technology handles execution but human provides strategic context.

00:13:32: That transition from being technical executor to a strategic custodian is vital.

00:13:37: however All of these ambitious plans, like deploying autonomous agents capturing runtime evidence rewiring the operating model.

00:13:45: They collide with a very harsh physical reality

00:13:48: servers The physical infrastructure required to run these dreams is becoming the ultimate bottleneck.

00:13:53: Oh

00:13:53: definitely So Rashidi and Sergio Vittal highlighted a macro level business story that beautifully illustrates this sheer desperation in the compute market.

00:14:01: right now What happened?

00:14:03: so the AI startup antropic recently saw massive eighty times year over your growth.

00:14:08: But they hit a wall.

00:14:09: They simply could not source enough processing power to sustain their models.

00:14:13: Wow, ADX growth is insane.

00:14:14: So what did they do?

00:14:15: To solve it...they brokered a massive six billion dollar deal to rent space and secure roughly two hundred twenty thousand GPUs at the Colossus Data Center.

00:14:25: Wait!

00:14:25: The Colossuses data center isn't that owned by Elon Musk's company XAI?

00:14:30: That's the punchline.

00:14:31: At the time this deal was happening, Musk was actively in court suing open AI and had previously gone on record calling Anthropic Evil.

00:14:39: And

00:14:39: yet Anthropic was so desperate for compute that they agreed to make their fiercest ideological critic Their landlord.

00:14:46: It's The ultimate proof That In current AI landscape Ideology takes a back seat To infrastructure.

00:14:52: As Sol Rashivi points out, compute is the new oil.

00:14:55: Right!

00:14:55: Having the smartest most mathematically elegant LLM algorithm in a research paper... ...is completely useless if you don't have electricity and silicon to run it at scale.

00:15:04: The ultimate winners of this era are entities that control power grid, physical floor space & chips

00:15:10: And that macro level scarcity of compute trickles all down into daily operating budget for normal enterprises.

00:15:16: For the last two years, we've kind of lived in this honeymoon phase where tech giants heavily subsidize the cost of AI to drive adoption.

00:15:23: Yeah!

00:15:23: We got used to paying a flat twenty-dollar monthly fee for unlimited magic...

00:15:27: Exactly but Tony Hinckley highlighted a major canary and the coal mine Github's recent move to token based billing for co-pilot.

00:15:35: That shift from a flat sauce fee To consumption base Token Billing is gonna catch A LOT OF CFOs off guard.

00:15:43: The underlying cost of operating these models is simply too high for vendors to swallow anymore.

00:15:48: Now every single prompt you write and every word the model generates in response has a literal microtransaction price tag attached

00:15:57: And the financial danger there is exponential when you tie it back to the autonomous agents we discussed earlier.

00:16:02: Oh yeah, if a human is typing prompts There's an natural speed limit to how much money they can spend.

00:16:07: But If You Have An Autonomous Agent Executing Workflows In The Background and That Agent Gets Confused & Stuck in A Retry Loop

00:16:13: It Could Automatically Fire Off Thousands Of API Calls A Minute.

00:16:17: Yes!

00:16:17: A poorly architected agent could literally bankrupt an entire quarterly cloud budget over long weekend

00:16:23: which is a terrifying reality, but it's elevating discipline called FINOps for AI from a niche interest to core survival mechanism.

00:16:32: Yeah I'm at.

00:16:33: Singh pointed out that platforms like the Unity AI Gateway are gaining massive traction specifically because they offer strict spend controls.

00:16:41: Enterprises are desperately trying build architectural circuit breakers.

00:16:45: They need single dashboard visualized token spent in real time set hard budget caps for specific agent workflows, and just automatically kill any process that starts hyper-consuming compute.

00:16:55: We have so rapidly transitioned from treating AI as this fascinating science experiment to realizing it requires the same rigorous metering governance an architectural planning As a municipal power grid.

00:17:07: It really is core infrastructure now From the non negotiable runtime evidence required by the EU AI Act To engineering multi layered memory systems to confronting the brutal new economics of token billing.

00:17:18: The signal cutting through all the noise on LinkedIn right now is that organizations that succeed won't be ones with the flashiest chatbots.

00:17:26: It will master the unglamorous foundational work of governance, architecture and fine ops.

00:17:32: Perfectly said!

00:17:33: If you enjoyed this episode new episodes drop every two weeks.

00:17:36: Also check out our other editions on ICT & Tech digital products and services cloud sustainability in green ICT defense tech and healthtech.

00:17:45: And before we wrap up entirely, I want to leave you with a concept to chew on as you look at your own enterprise architectures.

00:17:52: Both Dr.

00:17:52: Annette Doms and Omar Elmam Luke touched upon the profound limitation of these systems.

00:17:58: What's that?

00:17:58: We've spent this entire deep dive discussing how AI can perfectly recognize data patterns relentlessly optimize code and scale productivity to level humans just cant match.

00:18:09: but AI fundamentally cannot understand attachment.

00:18:12: Right it doesn't feel anything

00:18:13: Exactly.

00:18:13: It does not feel a sense of responsibility for the outcomes it generates, and it cannot forge human connection.

00:18:20: so as you go back to your desk to build the incredibly efficient perfectly optimized agent-driven architectures of tomorrow how will you ensure your organization doesn't become emotionally and creatively sterile in the pursuit of ultimate efficiency?

00:18:35: That right there is The Defining Leadership Challenge.

00:18:38: Thank you for joining us on this depth.

00:18:40: dive into the source material.

00:18:42: Don't forget to hit subscribe.

00:18:43: so don't miss our next analysis, and we will catch you next time!