DEV Community: Janet Hase Naranjo López

I have a new substack post about Nvidia drama... if you want to read 5 minutes of business/tech analysis from a real engineer go ahead please: https://janethase.substack.com/p/the-current-state-of-ai

Janet Hase Naranjo López — Fri, 13 Nov 2026 03:17:44 +0000

The Current State of AI - by Janet Hase

A not so brief text about important topics you should know about it...

janethase.substack.com

The Current State of AI

Janet Hase Naranjo López — Wed, 11 Nov 2026 03:37:13 +0000

A quick note on how this was made: I co-wrote this piece alongside Gemini using recursive chain of thought and custom voice constraints to synthesize the biggest industry shifts of 2026 into a single starting point for upcoming posts (still keeping hyphens to a minimum so I do not look like a bot XD). The curation, engineering lens, and perspective are are 100% human.

I was poking around recent industry reports lately to get a real pulse on where artificial intelligence actually stands as we wrap up 2026, and it turns out that the shiny “innovation lab” phase is officially dead. (like the Pokemons, just the shiny ones, not any dead ones).

I mean, for the last couple of years, everyone was obsessed with conversational demos and seeing what a chatbot could write in a browser tab. Today, the conversation has shifted entirely towards infrastructure economics, production bottlenecks, and some pretty wild corporate course corrections. Since I want to unpack several of these topics in future posts, I put together this news style roundup of the most relevant situations happening across the AI landscape right now.

1) The Agentic Shift (and the “Agentic Drift” Backlash)

The biggest architectural headline of 2026 is the transition from conversational chatbots to autonomous AI agents. According to McKinsey’s latest global survey, around 40% of large enterprises are now scaling agentic systems inside IT, knowledge management, and software engineering to execute multi step workflows rather than just answering prompts1.

Of course, what the marketing decks leave out is the operational toil that comes right after deployment. We are currently seeing a massive wave of concern around “agentic drift” (basically, what happens when an autonomous agent loses context mid workflow, starts hallucinating tool calls, or goes rogue in production) and the awkward legal question of who holds the pager when an automated workflow breaks. Even more interesting is that several companies that aggressively cut headcount expecting AI agents to run the show have quietly started rehiring human staff after realizing that broken data pipelines, governance overhead, and token costs sometimes make a human team cheaper and way more reliable than a digital workforce!

2) Coding Agents and the “Build vs. Buy” Plot Twist

One of the most surprising shifts this year is happening in enterprise software procurement. Because coding agents and repository level intelligence have matured so much, nearly a third of surveyed organizations reported cancelling plans to buy an off the shelf SaaS tool or feature because their own engineers simply built the functionality in house using agentic coding tools. Instead of paying bloated SaaS subscriptions for a simple internal dashboard or a connector, teams are spinning up custom internal tools in days, which is starting to put real pressure on traditional B2B software vendors.

3) The Great Inference Race: Edge AI and Custom Silicon

Anyway, going back to the point of infrastructure: running massive generalist models in the cloud for every tiny task is financially unsustainable, and the hardware market is finally reacting. On one hand, we are seeing a massive push toward specialized inference silicon (like Cerebras deploying wafer scale inference engines in AWS data centers and d-Matrix putting its Corsair platform into full production)2 to slash response latency from twenty seconds down to two while cramming five times more token capacity into the same hardware footprint.

On the other hand, Edge AI and small, domain specific models are stealing the spotlight. Instead of sending every API call to a massive data center, companies are deploying compact models (like Liquid AI’s LFM2.5 architecture running natively on AMD, Intel, and Qualcomm chips) directly onto laptops, vehicles, and IoT sensors3. It cuts cloud inference bills, eliminates network latency, and keeps sensitive data local.

4) The Crisis of Trust and the Saturation of Synthetic Noise

If the infrastructure layer is maturing, the public internet is dealing with the exact opposite problem. Synthetic media now accounts for a massive chunk of visual feeds and professional network posts, leading to what researchers are calling a “crisis of trust.” Users have developed rapid pattern recognition for automated slop, generic phrasing, and deepfake identity fraud, causing widespread audience fatigue. From an engineering standpoint, the recursive degradation problem is no longer theoretical: as the web fills up with unverified synthetic exhaust, teams are realizing that scraping the public internet to train the next generation of models is yielding diminishing returns, making verifiable, human generated data the most expensive asset in the room.

5) Governance Moves from Slide Decks to Calendar Deadlines

AI regulation also stopped being a philosophical debate and became an engineering compliance ticket. With the EU AI Act timelines locking into place alongside a patchwork of state level laws in the US, “tech sovereignty” and auditability are now deciding what actually ships to production. Leaders are no longer asking for vague ethics whitepapers; they want deterministic guardrails, data provenance, and clear audit trails before an AI system is allowed anywhere near customer data. It looks like a plain “we do not know how the AI leads to this result” is gonna be the answer anymore.

6) The Business & Geopolitical Single Point of Failure (Yes, the Nvidia Situation)

Finally, we cannot talk about the business side of AI right now without mentioning the elephant in the room: the recent scandal around Nvidia CEO Jensen Huang. Leaving the tabloid and paparazzi drama aside (because honestly, who cares about the gossip columns when we are talking about global compute infrastructure :D), the actual engineering and business story behind the headlines is a massive wake up call about supply chain vulnerability and “key person” risk.

When you strip away the sensationalism, the real issue that wiped nearly $190 billion off Nvidia’s market cap in a single day and triggered board, FBI, and shareholder investigations comes down to hardware provenance and export controls. The core investigation centers on Huang’s ties to an executive at Megaspeed, a Singapore based cloud company that bought nearly $2 billion in advanced Nvidia GPUs (yes, the newest chips available) and is now under heavy scrutiny for allegedly acting as a backchannel to divert sanctioned AI chips into China. Basically, the entire global AI buildout is so heavily centralized around a single hardware vendor that one executive’s personal blind spot (and a questionable cloud partner deployment) immediately turns into a national security investigation, a congressional briefing, and a trillion dollar market tremor. In SRE terms, the global AI economy built its entire stack around a single point of failure, and we are watching what happens when that node gets compromised.

Where we go from here

Basically, the current state of AI is a tug of war between massive capability and operational reality. Whether we are looking at autonomous code, cloud token budgets, or the geopolitics of the GPU supply chain, we are moving away from treating AI like magic and starting to see the very human, very fragile infrastructure underneath it.

Over the next few posts, I want to zoom in on a few of these specific threads (especially that SaaS “build vs. buy” shift, what “agentic drift” looks like from an SRE perspective, and how concentrated our hardware stack really is :D). Which of these six situations do you think we should unpack first for the next deep dive?

Subscribe to my substack

Process note:

Ingredients: 100% organic human thoughts and engineering rants. Packaged in a facility that also processes LLMs, so this post may contain traces of AI assisted brainstorming and structural iteration.

The useless use of AI

Janet Hase Naranjo López — Sat, 07 Nov 2026 02:07:28 +0000

A quick note on how this was made: Yes, I see the irony of using AI to help draft a post about the useless use of AI :D. However, instead of lazy one shot prompting, I co-wrote this piece alongside Gemini using recursive chain of thought and custom voice constraints to iterate on my rough thoughts and refine the structure (including a strict prompt rule to ban hyphens so I do not look like a bot XD). Intentional collaboration is the exact opposite of what I am complaining about below, and the core arguments, frustration, and perspective are 100% human.

I was looking at how people use AI in their daily workflows lately, and it turns out that a huge chunk of our token expenditure is going toward tasks that do not actually save any time (or worse, tasks that had no business being automated in the first place).

I mean, leaving aside the funny side of AI generation (like absurd memes, cursed images, or just poking around with prompts for fun :D), there is a growing trend of running inference on things that are simply not worth the effort, or that become completely meaningless once you strip the human out of them.

What I mean by “useless” (and what I do not)

Of course, I am not talking about using AI to automate repetitive infrastructure work, parse massive logs, or speed up a real technical bottleneck. That is what automation is actually for! What I am talking about is the performative, low value use of generative models where the tool adds zero net value to the world.

Basically, if you look closely at how people interact with LLMs and image generators every day, this useless usage usually falls into two buckets:

1) The Prompt Engineering Overhead: Have you ever watched someone spend fifteen minutes tweaking a prompt, on a prompt per prompt basis, just to write a three line email, a casual Slack update, or a happy birthday message? It turns out that by the time you finish giving the model context, regenerating the output because it sounds too robotic, and deleting words like “delve” or “testament,” you have spent triple the energy it would have taken to just type what you actually meant. We are turning a 45 second human interaction into a debugging session.

2) The Zero Sum Inference Loop: Person A writes three rough bullet points and asks an LLM to turn them into a five paragraph, overly polite corporate email. Then, Person B receives that five paragraph email, refuses to read all that fluff, and asks their LLM to summarize it back into three bullet points. Nobody gained any new context, no real communication happened, and we just burned compute and tokens to maintain the statu quo of corporate politeness. Anyway, going back to the point:

The cost to human creators

If this were just about people wasting their own time in a chat window, it would merely be ironic. However, when we apply this same effort avoidance mindset to creative work, the collateral damage hits actual human creators.

Again, leaving the memes aside (and the cute AI generated cats), the real problem starts when people use AI to generate “good enough” creative filler. Think of blog headers, indie book covers, event flyers, newsletter illustrations, or background tracks generated by a model just because someone does not want to pay an artist or spend time curating something real. In my opinion, this creates three massive problems for the creative ecosystem:

1) It kills the entry level commission:
Historically, small commercial gigs (designing a poster for a local event, illustrating a single article, or writing short copy) were the exact stepping stones that allowed emerging artists and writers to pay rent while developing their voice. When people treat those small tasks as friction to be automated away with a single prompt, they pull up the ladder for working creators.

2) It mistakes the output for the art:
There is a huge misconception that art is just the final image file or the final block of text. It is not. What makes a piece of writing or an illustration resonate is the human context behind it, including the lived experience, the intentional constraints, the happy accidents, and the specific point of view of the person who made it. When you replace that with a statistical average of scraped training data just to fill empty space on a page, you are not making art. You are just generating wallpaper.

3) It floods discovery channels with noise:
Because generating this kind of filler takes almost zero effort, it floods the exact same platforms where human creators are trying to share their work. It becomes harder and harder for intentional, human made craft to be found when it is buried under an avalanche of automated content that not even its prompter cared enough to make by hand.

The takeaway

For decades, the whole promise of technology and automation was that machines would handle the boring, soul crushing drudgery so that humans would finally have the free time to write, paint, think, and connect with each other. Instead, we somehow flipped the script: we are still doing the dishes and dealing with bureaucratic admin, while outsourcing our personal messages, our rough drafts, and our visual culture to a model.

Basically, just because we can throw AI at a task does not mean we should. Using AI intentionally means knowing where automation actually solves a hard problem, and knowing where the human effort, the imperfection, and the creator’s craft are the whole point. If we outsource our own voice and bypass human artists just to save five minutes on things that were never supposed to be optimized... what are we even saving that time for?

Subscribe to my substack

Process note:

I co-wrote this piece alongside Gemini using recursive chain of thought and custom voice constraints to iterate on my rough thoughts and refine the structure (including a strict prompt rule to ban hyphens so I do not look like a bot XD). Intentional collaboration is the exact opposite of what I am complaining about below, and the core arguments, frustration, and perspective are 100% human.

Mr. November and the Myth of the Green Dashboard

Janet Hase Naranjo López — Tue, 27 Oct 2026 20:28:19 +0000

I was wondering how often we let a big green dashboard fool us into thinking our systems are actually healthy, I mean, it was not actually just wondering but I used this topic in a job interview last week. It turns out, sports analytics solved this exact same riddle years ago and IT specialists can -and should- learn a lot from how scouts separate the box score, the huge amount of saber metrics and even the Money Ball ideas from the game tape.

Take Derek Jeter -even if you are not a big fan of the Yankees-. If you evaluate his career strictly through modern defensive metrics like Ultimate Zone Rating (UZR) or Defensive Runs Saved (DRS), the data is brutal¹. Statistically, he grades out as one of the worst fielding shortstops in baseball history. If Jeter were a microservice, an automated auditing tool would flag him for immediate deprecation.

Yet, nobody in their right mind would cut him from a roster. Why? Because metrics struggle to quantify high-leverage execution. When the playoffs arrived and the pressure maxed out, Jeter became “Mr. November.” The box score said his range was a liability, per se, but the game tape showed the guy who hit a walk-off home run in Game 4 of the 2001 World Series and anchored five championship teams. Now, look at how we run IT operations. We have the exact same blind spot, and it cuts both ways.

1. The Legacy “Mr. November” in Your Stack

Every engineering org has a Jeter in their architecture. It is that clunky, monolithic service written eight years ago. It fails every modern DORA metric you throw at it: deployment frequency is low, automated test coverage is sketchy, and its p99 latency isn’t winning any awards and might be eating a big piece of your budget error pie.

On paper, it looks like a technical debt that needs to be rewritten immediately. But when peak day traffic hits and your shiny, stateless, Kubernetes pods² start thrashing under unexpected connection pooling issues, that ugly legacy monolith just sits there, quietly processing 100% of your checkout transactions without dropping a single packet. The metrics say it’s a liability; the game tape says it’s keeping the company in business.

2. The 2010 Chargers and the “All-Green” Outage

What happens when you have the exact opposite problem, I mean, when you optimize purely for the box score and ignore the tape or the old school scout six sense?

Look at the 2010 San Diego Chargers. They finished the regular season ranked #1 in the NFL in Total Offense Yards and #1 in Total Defense Yards -of course this one means the least amount because they just allowed 4,345-. By the two biggest macro-metrics in football, they were an unstoppable juggernaut. Yet, they didn’t even make the playoffs. Their special teams unit -the joint between offense and defense- was catastrophic, giving up blocked punts and kickoff returns - they gave up 528 yards which is almost 200 above the league average³- that cost them game after game. We see ships capsizing -pun intended- like this every week in IT as SRE.

Your infrastructure metrics are great: CPU and memory utilization across your nodes are sitting at a comfortable 40%. Your CI/CD pipeline metrics are awesome: 100% of builds are passing and deployed in under four minutes, no disruptions. Your Grafana dashboards are a sea of calming green.

Meanwhile, your users can’t log in because an IAM policy update silently broke a cross-region API gateway, or a third-party DNS provider is blackholing traffic from half the country. You won the infrastructure box score, but you lost the customer game tape.

Watching the Tape in IT

Collecting telemetry is easy or is supposed to be easy, you got the point. We can log every container restart, trace every database query, and alert on every minor spike. But measuring everything is not the same as understanding anything.

When we rely strictly on raw metrics without context, we make bad engineering decisions. We spend months refactoring stable legacy systems just to make a code-quality dashboard look prettier, or we celebrate 99.99% server uptime while the actual user point of view is very bad and our systems looks completely broken.

Metrics aren’t useless per se, they tell you where to look at. But they will never replace the operational context of watching how your systems and your engineers, actually behave when the game is on the line.

And that’s the whole point of this, you can even have someone who codes as the gods: superhuman speed and reliability, flawless scripts, elegant lines of code but, does this uber-coder behaves with the same performance during a red sea of alerts? Not everyone is capable to just perform on average during an outage -of any kind- and of course we also have metrics for that⁴, I know, the whole point of this post it’s to complain about them and now here we are… Jokes aside, even those metrics will not measure exactly what we want, because the metrics are just the mathematical representation of something we defined and even that they mean something inherently, we need to learn how to interpret every piece of data and WHEN to interpret in a certain way.

Data are not flat 2D figures that look nice in a dashboard (I mean, they do look really nice) data are vectors that moves across time and direction, which means that location is important, understanding that and make your team able to do the same are the most unseen tasks as SRE⁵ .

Thanks for reading. Leave a comment and subscribe!

Do you know that you can integrate RAG into Gemini?

Janet Hase Naranjo López — Sat, 17 Oct 2026 20:01:55 +0000

I was asking random AI questions to my AI (Gemini) just to understand how it works with the skill that integrates data from Google Workspaces (emails, notes, google docs) just because I was wondering if every request was actually looking into the documents in Gdrive (more token consumption) and also exploring the option to make Gemini write mimicking my style... it turns out that only when you activate the skill goes into your own-authored docs, emails, slides and so on to get an idea of how do you write, which is ok for token expenditure and gives you an idea of you can improve the inferences made with that skill (you do not need to ask everytime "write like me", it will make it every time if the skill is active) I mean, you can look for some offline documents you wrote an upload them into your g drive or even you can setup the tone of your AI writing curating your own documents (and maybe emails too :D).

Anyway, going back to the point, if you give access to your workspace to Gemini you can use it -at prompt per prompt basis- with a "user level RAG" to improve your inference results just need to ask to add context searching for information in the specific folder you want, this way the result will have contextual search based on the documents stored there.

Of course this is not a fully RAG feature, basically because a standard RAG means you have a vector DB with already processed data to be consumed by the LLM during the retrieval for inferencing process.

Yeih! Now I have some GPU to use

Janet Hase Naranjo López — Fri, 16 Oct 2026 01:03:56 +0000

Just happy that I was able to attach my GPU in Vs Code using Colab :D

Now I have 15 mighty GBs of GPU to do something with ML... I mean, is not too much but given the current conditions in which we are and how easy is to get GPU these days, a little bit is more than zero.