
So they're going to dedicate Copilot to serving ads?
Microsoft is asking investors to "temper" expectations for quick financial returns from Copilot amid efforts to convince customers that paying "substantial" sums each month is actually worth it. After trialing the use of Microsoft's GenAI in their workflow, testers told the Wall Street Journal that they had mixed feelings …
copilot> What do you want to buy today?
user> Nothing, I have everything thanks
copilot> Are you sure? You are missing out on so many things that people in your area buy.
user> Like what?
copilot> When did you last time had a good night's sleep?
user> None of your business
copilot> You are staying up late and still wake up early. That's 4-5 hours of sleep?
user> So?
copilot> You need a carefully designed memory foam pillow to get the most out of the time you have for sleep.
user> My pillow is fine, thanks
copilot> Your typing rate is slow in the morning and you searched for pain killers.
user> None of your business
copilot> Just take a look at that link <amazon affiliate link to memory foam pillow>, won't hurt. See reviews.
user> Are they fake like yourself?
copilot> I understand that you are sceptical, but the pillow has 100% money back guarantee. I think it will help you.
user> I don't have time for this right now
copilot> I can place the order for you.
user> Ok do it, whatever.
copilot> Done. It will be delivered tomorrow.
*5 minutes later*
copilot> Have you heard of the benefits of CBD oil?
Tried it today - turning 12000 words of random notes into a PowerPoint deck. It basically just cut-and-pasted it into a pretty shit presentation. Not even particularly well formatted (not quite into additional limb or a missing finger territory).
Then asked it to compose an email thanking my co-worker for enrolling me in the trial. Was so fawning and gushing I actually blushed.
Deeply underwhelmed by Co-Pilot for office.
Yep. Our email is Exchange in the cloud, all our documents are in Sharepoint in the cloud, and we're migrating all our code to Github. MS pushing Copilot to Windows so it can scrape the remaining stuff off PCs is just closing the circle.
It's a US multinational so no CxO cares anyway, but any business outside of the US with a board which allows the same thing to happen is incompetent.
Yes. But to a bystander it would appear amazing that it could do it at all.
I had this experience way back when the first automatic translators appeared online. One was named Babelfish, which was a good start.
I was impressed. I showed it to a colleague, a French national of Lebanese origin. I got it to translate a randomly selected page off the BBC news web site into French. Isn't that amazing, I gushed.
No, he replied, it's shit.
But, interest piqued he asked to try it in the other direction. We got it to translate a page from some French news site (I forget which) into English.
Wow, that's amazing, he opined.
No, it's shit, was my judgment.
-A.
Whenever I see a claim that adding a given tools makes users of X, Y, Z more productive
... it makes me wonder about those users.
Take Copilot as a "coding assistant". Enthusiasts claim it makes writing code far more productive. What's their metric? Presumably something like volume of source code generated, or number of function points implemented, or time to completion for a small programming task (because LLM code generators haven't existed long enough to complete a large programming task).
But writing code isn't programming, and programming isn't software development, so those metrics aren't particularly useful. If a code-generation tool makes that much of a difference to your productivity as a programmer, then you're not a very good programmer, and you're certainly not a good software developer.
Also, a majority of software development is maintaining existing code: fixing, enhancing, or modernizing existing code bases. Often those are very large corpora — much too large to fit in the context window of any commercially-available LLM. So coding assistants won't have critical information for code-maintenance tasks.
It's harder to mount specific arguments about the use of LLMs for generating prose (or, god help us, verse), because natural-language discourse is far more subjective. But as someone with degrees in literature and writing, and who has taught writing and done a variety of professional writing, I've yet to see LLM output that I'd consider more than functionally adequate.
More importantly, using a tool to generate discourse is, I'd argue, fundamentally a cognitive trap. Writing is an important exercise in thinking: in organizing information, analyzing its structure and content, formulating arguments, making pragmatic1 choices, understanding audience, and so forth. Using an LLM sacrifices a critical opportunity to refine your thinking. That's counter-productive.
In short, the LLM proponents are measuring the wrong things.
1"Pragmatic" in the technical linguistic sense: being aware of and making an informed choice among various options for expressing an idea in natural language.
Do I detect the beginnings of an admission that this stuff has been hyped up out of all proportion, driven by the vast sums paid by big tech companies hoping to get in on the hype, and now that people have started paying for it and using it and discovering that it really can't do anything more than spit out stuff that you could have found just as easily with a simple web search, the realisation may be dawning that it might have been rather oversold?
Not to mention...
Employee "Here's the work you asked me to do boss".
Boss "Ah good. Did you use AI much?"
Employee "Yep. It was really good and did most of it! Saved me hours".
Boss "Then what the effing hell am I paying you for? Your fired!"
The MD decided that he'd like to try out Co-Pilot. 4 hours later an email was sent by his secretary banning Co-Pilot and also banning spending one penny with MS from now on.
We were already weaning thew users off of Office and Exchange and what it did to his PC was the last straw.
This was when it was spewing out pron.
We are going Linux and MacOS from now on. MS is dead to us.
Carry on forcing people to take your half baked tech when they don't want it. It is a great way to get rid of the dross from your customer base. Eventually, a few very large companies will do the same. How's that bonus then SatNad?
I suspect that the experience you describe will be the exception rather than the rule.
Microsoft know how most large company bosses think, and they've already identified that touting trivial productivity gains will be sufficient to persuade them. They'll keep on shovelling out "research" that "proves" people are more productive with Crapalot, and in the grand tradition of the Emperor's New Clothes, the bigwigs will suspend their disbelief and insist that it is rolled out. In addition to ENC, as soon as famous consultancies start bragging about the massive benefits realised at a competitor, that will tap straight into the executive FOMO.
Like a tsunami of sewage, AI is rolling in unstoppably.
I wonder what goes on with so many organisations switching to Teams and Microshaft ecosystem.
There is always universal backlash from the users and push back from the top.
I am yet to figure out what value add Microshaft brings, but procurement seems to be happy with it.
For the lawyers, these are all assumptions.
They have a vast infrastructure dedicated to security - because people tend to forget they're actually the reason you need that to start with.
They have ruined usability - so they can prove to shareholders that it doesn't matter how bad they make it, people are locked in (and they can always revert this, claiming 'improvement', although anyone looking back at the promises of more productivity will have to be convinced to ignore the many lost manhours after they changed the user interface once again - and they're at it again with Teams and Outlook).
They have deep data analytics so you can stop people from copying confidential information (unless they use a camera and make pics of screens) - because they need that anyway to have a look at all that juicy IP they could maybe use themselves. In that context, the need for CoPilot to 'learn' means allowing it to trawl through everything..
Do. not. trust. them.
Who is buying this bullshit? Who is it really saving time for so much so that it is 'indispensable'. I don't for a minute believe that everyone all of a sudden needs copilot. They'd love you to think they do because they've put so much time and effort into it, oh and lots of cash too.
They can't let it fail because they've bet the house on AI. Only it is going to, because it's worth nothing beyond being a shiny toy for people to play with. I would argue it would probably cost more productivity in that respect if companies were daft enough to enable it.
> “… and people have very high expectations," he told an audience of investors
I think what he was saying that people want a good product and not just some semi finished crap bolted on openai’s api’s.
Also, the claim it saves 10 hrs a week is baseless. I’m using these various tools daily and their effectiveness is marginal, akin to being introduced to a typeahead in the search.
There is potential for sure, discounting all of the hype, but it will be acieved by some other company capable of building better products.
We have invested heavily and soon we will need some greater fools to buy us out. We tried it and it works great, honest. We thoroughly recommend it ... to our competitors. If your IT people are trash talking it that must be because they are afraid it will replace them. If it did not rake in cash at your friend's company that is either because he is doing it wrong or he should have bought the enterprise version.
Now how about investing in some sure fire high return sub-prime loans or block chain tulips?
I'm a Linux & Mac user, and definitely not a fan of Microsoft. But I've been using Github copilot recently with some great results. Last week I was writing a Lightningjs app for a smart TV. I've never used Lightningjs before and knew nothing at all about it (or indeed Javascript). Github copilot can be scarily good at times. Even with a language and framework that I had zero knowledge of on Monday, by Wednesday I had a basic but working home automation app running on a TV, able to control lights and other home devices. Copilot was able to guess what I was trying to do and would suggest code to do it. It wasn't always correct, but sometimes it was simply astonishing and really helped me to learn rapidly.
Right, but that's really just showing you the power of the LLM when it can scrape through everybody's code. Or did it tell you where it got which bits from? Doesn't really matter, the domain-specific stuff like programming should make the literate dream a possibility. But, considering there are now open source models out there that can do just as well, how much are you prepared to pay to use it?
> I've never used Lightningjs before and knew nothing at all about it (or indeed Javascript). Github copilot can be scarily good at times.
If you know nothing about either the library that you somehow decided you needed, nor even the programming language that somehow you decided was appropriate, just how would you judge how good it is?
-A.
All you high-falutin' business execs who swear on Sun Tzu, explain to me why :
1) you are willing to not only give your internal company workings to a competitor (Borkzilla is nobody's friend), but are willing to pay for the privilege
2) you have apparently no qualms to hand over such data to a company you cannot trust to either not benefit from it, or not sell that data or use it somewhere else
Is it because, when Borkzilla starts a new branch that does what your company does, you're hoping to get hired as branch manager ?
Good luck with that.
If it were real, most office workers would use an X percent improvement in "productivity" (however measured) to goof off by that amount. I know I would.
My value to my employer lies in what I am able to do, not (within reason) in how long it takes. I rather hope that most of the rest of you are in that position.
-A.
I had a little play out of curiosity. Most of the responses felt like they were directly lifted and reformatted from search engine data / wiki etc. I had a go at telling it to write Z80 code and it spew out loads of irrelevant stuff until I got a bit of code (hello world, yellow text on a brown background). after that I went to bed and didn't bother any more.
This "No generating real people" thing is definitely going to become dated and unfashionable. You can download Stable Diffusion and generate pictures of real people easily. The models are out there and almost entirely believable, especially for laypeople not looking out for AI images. The genie is out of the bottle, but they just want to cover themselves in case someone says they used Copilot to generate a pic of Princess Kate trying to leave the country.
You can delegate copilot to attend teams meetings on you behalf. So it joins the meeting and captures a transcript etc. I was at an event the other day and this customer was talking about their experience around doing this whilst on the co-pilot pilot. They have a morning status meeting which is very boring. So, one day, one person sent co-pilot to 'represent' them. Next day, two people sent co-pilot. And so on. They were sayiong that pretty soon the morning status meeting would be entirely populated by co-pilots, rather than people.
Of course, this is when the co-pilots will start talking to each other. Otherwise known as the singularity :)
So the business case for Copilot is proving that some (many? most?) meetings are pointless and should be cancelled?
That explains the "10 hours a week" claim!
Although once you've cut back your meetings to the ones that actually matter, why would you keep paying MS to attend them?
I'm sure there are already cases of email conversations between LLM agents. And we'll see more and more of this over time. Machines just generating prose that's never seen by any human and has no useful effect whatsoever. Just consuming resources.
We already have more or less constant wars between bot armies that are no longer being actively supervised by anyone and automated network defenses; this is simply a milder version of the same thing. Automate complex systems enough and eventually they'll simply run on their own to no productive end, as long as you keep them powered on.
Set it to exploiting vulnerabilities in smart contracts and mixing the proceeds. Have it create MEV bots and front-run transactions. Pump-and-dump cryptocurrencies, or create new shitcoins and DeFi projects and then rug-pull.
Use it to automate help-desk scams, phishing, BEC attacks.
Automate conventional cons: Ponzi schemes and other bogus investment vehicles, romance schemes, 419s, grandchild-in-peril.
None of that would be difficult; the only problem is getting in before you're drowned out in the noise of every other competent criminal enterprise doing the same.
Oh, make money by providing actual value? Hmm, that's a bit tougher.