Selected AI Topics (August 2025)

This blog shares some links about some interesting AI topics. Along with some of my thoughts of course. Including about where AI might be headed. Note that this is based on what I read from various sources — and I lack close ties to AI practitioners. You may be in the same situation: trying to keep up with AI trends and what’s useful/interesting, while doing $DAY_JOB. Or in my case, acting like I’m retired at least some of the time!

Some of the topics that follow:

• What is the Overall AI Game plan (by the big vendors)

• Is “bigger LLM’s are better” true? It certainly seems to be a popular mindset.

• Somewhat related (in terms of your definition of “better”): overall performance, cost, and ROI are increasingly appearing in discussions. About time!

• ROI on ML vs AI

• ML and AIOps FTW?

• Love that Agentic AI and MCP

• Applications of AI

• Agentic AI links (update!)

FWIW, I tried several AI agents to generate an image, got some amusingly bad results, and after specifying very precisely I got the following out of Gemini. Note that it (a) gets the text labels right, and (b) vastly exceeds my artistic skills!

What is the Overall AI Plan?

I haven’t seen a good explicit explanation of WHY the big AI companies are building out all the huge costly datacenter spaces. Apparently they expect a huge services business, either with companies and end users making paid use of their giant LLM’s, or buying AI compute etc. services for their own LLM’s, or tweaking parameters / doing different training in their own copy of a vendor’s LLM.

A lot of the discussion still seems to be of the “my LLM has more nodes or parameters” than the last one. Bigger is better? Is that just because that metric is simple enough to be useful in marketing? Or is it because of labor costs of carefully curated training data, multiplied by having many specialized LLM’s?

I personally suspect ultimately value (and greater accuracy?0 will be via specialized AI models for specific topic areas / use cases. Developing such a model will likely require investment in training an AI model based on curated data, which has high human labor cost.

So is the expectation that development and scalable runtime delivery of the services of such a model will leverage the giant AI company platforms? Will require such platforms, versus running e.g. in your laptop or smartphone?

One hopes the big AI firms got the growth potential right and this is not another California gold rush! The huge cost of all the buildouts — the “AI arms race” — is it based on data or hype or what? If the demand isn’t there, could there be enough investment money at stake that the consequences to the economy and banking could be really ugly?

This blog continue with this topic including some interesting links, in the next section below.

The rest of this blog considers some alternative perspectives.

Bigger is Better, Isn’t It?

Let’s revisit the LLM model size topic in more depth now…

I just saw a vendor claim of a yet another huge-number-of-parameters-model with some optimizations around training. And found myself thinking “why is consuming massive amounts of costly resources a Good Thing? Isn’t there an alternative?”

Is this about a bragging rights / attracting funding and marketing based on a single metric (size of model), etc.? If so, why, what am I missing? Why do the big vendors appear to be fixated on this, is it ease of competitive marketing? Or just that they not talking about other directions they are pursuing?

A Nvidia technical article tackles SLM (small LM) vs LLM directly. It is titled “Small Language Models are the Future of Agentic AI”, https://arxiv.org/pdf/2506.02153. The article is fairly short and quite readable. I did includee that link in a prior AI posting, AI Quick Takes and Musings. If you haven’t read it, highly recommended!

See also Hype, Sustainability, and the Price of the Bigger-is-Better Paradigm in AI,  https://arxiv.org/abs/2409.14160 for some related points. Good stuff to consider!

Related links and blogs providing two persons’ commentary:

• https://www.linkedin.com/posts/jalovell_agenticai-slm-llm-activity-7343306626797346816-DszR/?utm_source=share

• https://x.com/RubenHssd/status/1952048085133369751

The latter of those two posts includes some interesting points, including the following (my rewording):

• Central LLM investment vs. market for API services: claims 10:1 ratio.

• The assumption that centralized generalized LLMs are key may be flawed, and is about to get very expensive.

• For narrow repetitive tasks, huge models are overkill.

• Small language model costs 10-30 time less than huge one.

• And can be fine-tuned much more quickly.

• And can run on the edge, including laptop or phone.

• Smaller models can out-perform larege ones.

It continues into agentic approaches, and other topics of interest.

Another link in my prior AI Quick Takes blog notes that if an LLM is too big, it in effect memorizes the training material — so as the LLM scale goes up, you need a MUCH larger training data set to avoid “memorization”. See https://www.linkedin.com/posts/pascalbiese_how-much-do-language-models-memorize-ugcPost-7340076759083577344-LodU/?utm_source=share

Yet another link I’ve posted previously suggests that having a human involved decomposing the stages of solving a problem may perform better: an Large Reasoning Model given too big a problem to solve gets lost in effect deeply analyzing unlikely possibilities: https://ml-site.cdn-apple.com/papers/the-illusion-of-thinking.pdf. (I think I had and posted another link along this line, but if so, I can’t find it. Oh well!)

The general conclusion: too big a model in relation to the amount of training data is counter-productive! So big is not necessarily better. ROI, and ML versus AI

Keith Townsend recently posted an interesting perspective, that Machine Learning not LLM is the “real business powerhouse in 2025” with “repeatable ROI”. https://x.com/CTOAdvisor/status/1944708699651555557.

I tend to agree. It appears that targetted ML offers faster paths to value. LLM use has a bigger learning curve, e.g. learning prompting with pre-priming (example results), task decomposition into sub-tasks, etc. A lot of skills-building, only worth it for repeated use? What I’ve seen looks like working with a very smart but inexperienced assistant to get results. Not instant — it’s a skill to be acquired!

Keith also notes that LLM’s shine at natural language tasks. Perhaps obvious but worth calling out! I currently wonder if that means they’ll be a front end for more task-oriented LLM agents.

So if you’re looking to use AI to replace call agents etc., speech to text conversion fed to an AI to answer a question or initiate an action, that (to me) makes good sense, fits what current LLM’s might do pretty well. And if you are a business looking to solve such a problem, maybe you don’t want to be in the LLM business, you want to leverage them.

Keith notes the drawbacks to LLM’s as well.

So leveraging LLM’s as a Service (LLMaaS?) is perhaps what the AI giants are hoping you’ll do.

Concerning ML, one great application appears to be MLOps.

The pre-requisite is arguably having vast (network, application, storage) data feeds populating very large databases.

Obviously tapping into various data feeds takes time and will need prioritization. And some thought to ROI: just storing data that you’re not going to act on could get costly. On the other hand, if you’re touching some performance data source, it may make sense to extract all you can (within reason), rather than having to go back later to get more data.

All this is especially true of going beyong measuring overall application response times to measuring micro-services and application component response times.

Companies that come to mind that do this sort of thing:

• Kentik (had the data, rapidly adding ML tools and alerting)

• Selector.ai (startup with data and ML skills and now experience)

• Splunk (big data, seems to be adding ML fast)

• No doubt others

Agentic AI

John Capobianco’s convinced me that using the LLM as a front end, with some awareness of MCP tool capabilities, has tons of usefulness. What particularly strikes me is the possibility of having some of the agents then be specialized LLM or other tools. The ROI seems here, and I like the thought that it might reduce the need for prompting wizardry skills.

His latest (as of this writing): https://www.linkedin.com/posts/john-capobianco-644a1515_gemini-cli-a-cli-for-the-age-of-ai-activity-7358914350851452932-pcMi/

This strikes me as a great avenue for exploiting general LLMs and reducing entry-level skills.

I’m going to move on, as Agentic AI will likely re-appear in future blogs.

Some interesting recent links about Agentic AI:

• Hank Preston of Cisco with a how-to article about setting up Ollama, pyATS, CML, etc. to do MCP: https://blogs.cisco.com/learning/creating-a-netai-playground-for-agentic-ai-experimentation

• Phil Gervasi (hi Phil!) of Kentik: Taking AI Apps From Prototype to Production

• John C post re Itential/Selector MCP discussion: https://www.linkedin.com/posts/john-capobianco-644a1515_had-an-awesome-discussion-with-peter-sprygada-activity-7348726104100143105-1DcH/

• And a Futuriom blog re Itential and MCP: https://www.futuriom.com/articles/news/itential-puts-ai-agents-and-mcp-to-work/2025/05

AI Apps

Where things are right now, we need to learn by sharing what works and what doesn’t work. So here are some links…

The following link provides great how-to advice on how to succeed at web coding via AI.

• https://www.linkedin.com/posts/gravette_ai-claudecode-codex-activity-7351995211603972098-bgsE/?utm_source=share

Here’s a post about what AI might be best suited for (and I read ROI into this):

• Dangerous, dull, dirty, difficult: AI apps/use cases: https://www.linkedin.com/posts/sol-rashidi-mba-a672291_futureofwork-workforce-ai-activity-7346534454359715843-GIDI/

And here’s one vendor that claims intelligent flash use can significantly reduce AI training costs:

• https://www.linkedin.com/posts/jackpoller_ai-artificialintelligence-machinelearning-activity-7354215567840419840-ri3W/

Conclusions

Be skeptical about ‘big LLM is better”. Sometimes it is worse (cost, quality of results, etc.)

Consider ROI (both development and operational costs).

Consider: what’s the low hanging fruit? What’s going to be hard to solve?

Look for success stories, reduce your learning time.

Share what works to help others.

Share how you went about learning, e.g. prompting skill for report or image generatiuon, and what works/doesn’t work.

Memo to self: look for more cases of pre-priming and prompting. (I’ve got two e-books on the topic that I’m reading/skimming through.)

I’m still mixed in my feelings about what I’ve seen, most of which resemble talking (or typing) instructions to a smart but very inexperienced helper through some task. That’s only a win if it lets you amortize the learning curve across multiple practical uses. For coding or report generation, it’s a skill one might wish to acquire. Is it useful if it is a skill you’re only occasionally going to use? (“Personal ROI”). I guess “yes”, if you enjoy doing such exploration.

Disclaimer

I’m not an AI researcher and can only pass along what I’ve seen and read (and thought I understood). In addition, it seems highly likely there is much good work out in the field but little publication about it, where competitive edge is at stake.

Miscellany

Reminder: you may want to check back on my articles on LinkedIn to review any comments or comment threads. They can be a quick way to have a discussion, correct me, or share you perspectives on technology.

Hashtags: #PeterWelcher #CCIE1773 #SLMvsLLM #AgenticAI #AIApps

FTC disclosure statement: https://www.linkedin.com/pulse/ftc-disclosure-statement-peter-welcher-y8wle/

Twitter: @pjwelcher

LinkedIn: Peter Welcher, https://www.linkedin.com/in/pjwelcher/

Mastodon: @pjwelcher@techfieldday.net

BlueSky: https://bsky.app/profile/pjwelcher.bsky.social