Category: AI

Posted on

Is China’s DeepSeek a Threat to Microsoft’s AI Strategy?

All the tech world continues to be abuzz as it woke yesterday morning to the release of the DeepSeek-R1 family of AI language models, funded and created in China. At one point during the day yesterday tech stocks had plummeted – only to return to more or less normalcy (whatever that means) today.

Has the AI bubble burst, as many (including me) have predicted? Is Microsoft’s AI strategy in jeopardy?

I don’t think so. But let’s dive into the details to fully understand its ramifications.  

Planning and Reasoning

DeepSeek rivals the most advanced AI models (such as OpenAI’s o1) in its ability to plan and reason in order to solve a problem.

Consider this problem. Notice that DeepSeek begins with a <think> step in which it plans its attack on the problem:

>>> what is 2*10 divided 4 with pi added to it?
<think>
First, I need to calculate the product of 2 and 10.
Next, I will divide that result by 4.

Finally, I’ll add the value of π (pi) to the quotient obtained from the division.
</think>

Then it executes its plan:

1. **Multiply 2 and 10:**
   2 \times 10 = 20

2. **Divide the result by 4:**
   20 \div 4 = 5

 3. **Add π (pi) to the quotient:**
  5 + \pi

Since π is approximately 3.1416, we have:
   5 + 3.1416 = 8.1416

**Final Answer:**
 {8.14}

At first blush it seems remarkable; DeepSeek’s methodical thinking process at least appears to closely mimic how humans think (very similar to how OpenAI’s o1 model functions).

But it’s worth pointing out that DeepSeek is not a from-scratch LLM – it’s a distillation of other foundation models (Meta’s Llama and Alibaba’s QWEN). Essentially DeepSeek was “taught” by them – meaning DeepSeek did not need the vast resources that their pedagogues required.

DeepSeek is Open-Source

Importantly, DeepSeek’s models – there are six of them varying by number of parameters – are freely downloadable; all but the largest fit comfortably on a decent-size desktop (the 14 billion parameter model occupies about 9GB on my machine).

Moreover, DeepSeek is offered under the permissive MIT License (“…without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software…” ) which means that developers can use it in applications pretty much any way they want. Purportedly, even the model weights are available, although I haven’t found them yet (DeepSeek’s GitHub repo, created yesterday, is here.) And finally, the DeepSeek team has published a comprehensive paper outlining their technical methodology – meaning that anyone, in theory, can reproduce their work (apparently there is work in progress to do just that).

Microsoft and DeepSeek (and Other LLMs Generally)

But how does it affect Microsoft and its strategy?

It’s no secret that LLMs in their short lifetime have become commoditized, a fact Microsoft has wisely recognized. Since ChatGPT was announced just a little over two years ago, dozens of foundation and frontier models have appeared; the LLM repo HuggingFace now offers well over 1 million fine-tuned large and small language models.

In fact, models form just one part of Microsoft’s overall strategy, which comprises a far more expansive and inclusive view of AI in the enterprise. For Microsoft, the true value of AI lies in the myriad applications it can power – and that developers can build using it.

Microsoft’s vision for the AI-powered enterprise includes providing user interfaces to LLMs connected to corporate data (Copilot); offering a wide assortment of LLMs for developers to make use of; deep set of AI-focused tools for developers to use (AI Builder, AI Foundry, and ML Studio); and lastly, providing access to the ”fuel” that powers enterprise AI applications, data, including productivity data (Microsoft 365), analytical data (Fabric), and corporate applications through connectors.

Political Ramifications of DeepSeek: Another TikTok?

It’s unlikely that the current US administration will, or can, block the use of DeepSeek as it nearly did with TikTok. By now DeepSeek has been downloaded to millions of computers (including, as you’ve seen, mine); blocking it as a purely technical matter will be close to impossible.

Nevertheless, because DeepSeek originates in China, geopolitics cannot be ignored. I asked it point-blank if China’s ruler Xi Jinping is a dictator; after an over-600-word dissertation, it replied (note that, as mentioned, I am using a downloaded version of DeepSeek; evidently the online version, hosted in China, is more circumspect):

Labeling Xi Jinping as a dictator depends on one’s perspective of China’s political system and the definition of dictatorship applied. Considering the unique governance structure and collective leadership within the CCP, it is complex to apply traditional Western definitions of dictatorship to China’s context.

I then asked Microsoft’s homegrown Phi4 model the same question and received more or less the same diplomatic, noncommittal answer:

Ultimately, whether one views Xi Jinping as a dictator may depend on their interpretation of political systems, definitions of democracy and authoritarianism, and perspectives on governance in different cultural contexts.

DeepSeek also (somewhat surprisingly) provided relatively objective answers on topics controversial in China, such as the 1989 Tiananmen Square massacre; however it refused to answer a question about the state of the Chinese housing market, saying it only provides “helpful and harmless responses.” Hmmm.

Microsoft, OpenAI, and DeepSeek

The interactions above raise some important questions. Although DeepSeek appears to have achieved a new level of LLM transparency, we do not yet know to what extent bias and harmful content are filtered or guardrails have been applied – whereas Microsoft and OpenAI scrupulously follow Responsible AI methodologies. Additionally, not much is known about the Chinese startup that created it, which might raise concerns about using DeepSeek in mission-critical applications.

It’s not perfect by any means. DeepSeek’s knowledge stops at July 2023 and it doesn’t appear to have scoured every available internet source (for example, it didn’t know that I worked at Microsoft so hadn’t seen either my personal website or LinkedIn).

Nevertheless, it’s possible, likely even, that DeepSeek models could show up in Azure’s stable of LLMs, to be evaluated, compared, tested, and perhaps deployed within applications on Azure. So at least in one way DeepSeek could complement Azure.

But, when asked how DeepSeek could impact Microsoft, the LLM itself had a slightly more ominous answer:

…in the fast-evolving tech landscape, companies like DeepSeek could potentially compete with Microsoft in areas such as AI-powered search engines, enterprise software solutions, or cloud services. For example, if DeepSeek develops advanced AI tools that rival Microsoft’s offerings (like Copilot for Office), it could influence market dynamics. Similarly, partnerships or collaborations between the two companies could also emerge in the future.

Well, I doubt either will happen; it’s hard to see a DeepSeek-based Copilot as it’s so tightly integrated into Microsoft 365. My view is rather that DeepSeek – just one component of the overall AI stack – will prove generally beneficial to the overall AI ecosystem and Microsoft in particular.

But we’ll see.

Posted on

Barry’s Holiday Wish List for Windows

Now as you all know, I love Microsoft Windows. I have used it and its predecessor DOS since the early 1980s (yes, I’m old); its evolution over the years has been little short of amazing. And of course I worked for Microsoft here in the Pacific Northwest for a decade and a half.

That said.

I have a number of everyday gripes that I just wish Microsoft would fix for once and for all. None of these are, in my view as a software developer of fifty years’ standing (wow) appear very difficult – so please, team, just fix them!

In no particular order:

Make Authenticator Work With Apple Watch

Unlike my younger comrades, my iPhone is not an appendage to my body. Often (real often) I’m at my PC and some account times out, I have to type in the magic number and…where did I leave my phone?

I imagine there’s some Bluetooth security issue with making it work on the watch, but why can’t we fix it?

Let Outlook and Teams Share Identities

How many times have you had to sign into your email account (using Authenticator) and moments later had to repeat the process with Teams?

This feels like the relevant engineering groups have to have a meeting. Just saying.

Settings and Control Panel

Just this morning I was attempting to move the default location of Windows Update from C: to D:. It’s not clear this is even possible, but searching for answers yields any number of inconsistent results – because with nearly every release of Windows some Settings move, change, are deleted, or move from Control Panel to Settings, or whatever.

Dear Microsoft: rid of Control Panel for once and for all. Or Settings. Whatever. And then don’t change the UI. Ever.

Sound: Part 1

Save the last volume setting and don’t reset it to the super-loud default for no apparent reason. Every time I load Teams or YouTube Windows blasts my ears.

Sound: Part 2

This one’s a bit esoteric but applies, I imagine to any musician attempting to use Windows. I play the pipe organ (badly) and use Hauptwerk. There’s a known issue with the Windows Audio subsystem in which input from MIDI keyboards is batched – which means when you press a key there’s noticeable latency. It makes Windows essentially unusable for MIDI (music) devices unless you buy external hardware (I use an inexpensive Presonus AudioBox).

This works with no issue on a Mac – should be easy to fix on Windows.

Clean Up the C: Drive

I’ve complained about this before. Microsoft installs apps on C:\Program Files, C:\Program Files (x86), C:\Users\You\AppData (and three folders within)…why? (And \AppData is hidden!)  Macs just have /Applications. It’s a mess.

Moreover: there’s so much junk on the C: drive, some of it from Microsoft, a lot of it from vendors – like the 13GB (!) installer for my keyboard and mouse from Razer. There are .DMP files, log files that never get purged or deleted but rather grow forever, literally occupying tens of gigabytes of space. Microsoft should develop and enforce rules about how the C: drive is used. It’s the Wild West now.

What Changed?

Because I have a relatively small C: drive (256GB SSD) I keep an eye on free space (I wrote my own df command-line app to report free space.)

One day I have 13GB free. Another 8GB. Then 4GB, 2GB. Then 10GB. Why? What changed? (It wasn’t a Windows Update.)

I use the invaluable Wiztree to report on disk usage but it doesn’t show what changed from one day to the next. And I would like to know – and control – when and where the downloads happen.

Why Is It Slow?

Recently on my machine (an i9 with 64GB RAM with up to date antivirus) that old reliable Ctrl-Alt-Del app Task Manager takes forever to load. And sometimes (like right now) it displays the white “(Not responding” title bar).

Why? Not even Bing Chat or ChatGPT can help, other than to give some banal and useless advice.

Ultimately I’d really like to know What My Machine Is Doing, and have the tools to (easily) dive down to the bit level. I fear, however, that’s a whole new OS rewritten from scratch.

Posted on

AI: What’s Wrong and How to Fix It

Want to know how generative AI works?

Imagine a newborn child. Now, just for fun, imagine that this child – we’ll call him Karl – is born with the ability to read. I know, no way, but suspend your disbelief for just a second.

So Karl can read. And by the way, he can read really, really fast.

Now, just for fun, let’s give poor Karl the entire contents of the internet to read. All of it.

Task done, everything Karl knows is from the internet.

Most infants learn basic, foundational things as they grow up. “Hey look, I’ve got hands! Oh wow, feet too! The dog’s got four legs… and a tail…and it barks!”

But Karl never learned these things. Karl only knows what he read on the internet. So if we ask Karl to write an RFP (Request for Proposal, a common business document) that’s like others our company has written, he’ll probably do a fantastic job. Why? Because he’s read zillions of them, knows what they look like, and can replicate the pattern.

However, Karl can’t get common-sense relationships, as Gary Marcus elegantly pointed out in this blog post. As he notes, Karl may know that Joe’s mother is Mary, but is unable to deduce from that fact that (therefore) Mary’s son is Joe.

Nor can Karl do math: ask him to calculate 105 divided by 7 and unless he finds that exact example somewhere in the vast corpus of the internet, he’ll get it wrong.

Worse, he’ll very authoritatively return that wrong answer to you.

That’s a loose analogy of how Large Language Models (LLMs) work. LLMs scrape huge quantities of data from the internet and apply statistics to analyze queries and return answers. It’s a ton of math…but it’s just math.

In generating an answer, LLMs like ChatGPT will typically create multiple possible responses and score them “adversarially” using mathematical and statistical algorithms. Does this look right? How about that? Which one’s better?” These answers, however, are tested against patterns it finds – where else? –  in the internet.

What’s missing, in this writer’s humble opinion, is an underlying, core set of common-sense relationships – ontologies to use the technical term. “A mammal is a lifeform that gives live birth and has hair. A dog is a form of animal. A horse is a form of animal. Dogs and horses have four legs and tails.” And so on.

LLMs need what is called a “ground truth” – a set of indisputable facts and relationships against which it can validate its responses, so that it can – the word “instinctively” comes to mind – know that the mother of a son is also the son’s mother.

Microsoft claims that Bing Chat leverages Bing’s internal “knowledge graph,” which is a set of facts – biographies of famous people, facts about cities and countries, and so on, and this is a start, for sure. More interestingly, Cycorp, which has been around for decades, has built enormous such knowledge bases. And there are undoubtedly others.

What I’m advocating is that such knowledge bases – facts, properties, relationships, maybe even other things (like Asimov’s Three Laws) underly LLMs. In the adversarial process of generating answers, such knowledge bases could, in theory, not only make LLMs more accurate and reliable but also – dare I say it – ethical.

(This post was inspired, in part, by this marvelous paper by the late Doug Lenat and Gary Marcus.)