That seams to be one of the best use cases for this, provided you can run it without giving Microsoft a lot of data directly. Else this is a no-no for enterprises imo.
Sure, but most (if not all?) of us have language in our customer agreements that Microsoft doesn't have rights to just start browsing our OneDrive libraries and Exchange inboxes without a support contract explicitly authorizing access for support purposes. Just like, sure, I technically have access to a bunch of PHI at my workplace - but touching it without a valid reason is a one-way ticket to the unemployment line.
But unlike Microsoft's other services, AI needs data to train on - are they going to be content just sourcing this from the internet at large or have we finally hit an inflection point where they try to get business and enterprise customers to give access in exchange for these features? That's a big no-no to anyone in a field that deals with sensitive information, and in most all cases an equally big no-no to anyone in a company that wants to keep things confidential.
> Sure, but most (if not all?) of us have language in our customer agreements that Microsoft doesn't have rights to just start browsing our OneDrive libraries and Exchange inboxes without a support contract explicitly authorizing access for support purposes.
You don't have language in your agreement that Microsoft cannot use your data to provide or improve service. You have language that says _support personnel_ cannot access your data without approval.
Microsoft already indexes your mailbox, OneDrive, etc.
Viva Topics has been around for ~2 years which mines content across various services.
Many large enterprises already store almost everything on 365 and azure. Other than budget, the only thing they need to turn on copilot is for it to adopt the same data privacy guarantees already available in 365.
This is a service for companies that already store all of their e-mail, documents, files, chat, etc with Microsoft. If there was a concern about data privacy that company wouldn’t be using Microsoft 365 in the first place. I think the bigger concern is how your data is used for training. If it is at all, I don’t know.
The integration mentions working with Microsoft Graph API, and it’s doing the generation on the client, so seems plausible this has no more access than any other OAuth app. The Graph API can be locked down with Access Policies. I’m sure there will be features to further cordon off data to the AI, similar to how OneDrive has an encrypted vault which is not accessible from the Graph API.
Rewind.ai does full time screen recording taking advantage of macos native OCR so you can do full text search of anything that has ever appeared on your screen - i'd expect they're working on throwing a language model on top
I think we'll get a glimpse of something similar in Windows 12, though I do wonder if they will limit it to Microsoft 365 users, since the compute costs of billions of users that only bought the Windows license would quickly add up.
The article does not provide a specific release date for Microsoft 365 Copilot, but it mentions that the company is testing it with a small group of customers to get feedback and improve their models as they scale. It also states that Microsoft plans to bring Copilot to all its productivity apps in the months ahead. However, pricing and licensing details have not been announced yet.
They've really put the summarisation and Whisper capabilities to good use!
"Unlock productivity. We all want to focus on the 20% of our work that really matters, but 80% of our time is consumed with busywork that bogs us down. Copilot lightens the load. From summarizing long email threads to quickly drafting suggested replies, Copilot in Outlook helps you clear your inbox in minutes, not hours. And every meeting is a productive meeting with Copilot in Teams. It can summarize key discussion points — including who said what and where people are aligned and where they disagree — and suggest action items, all in real time during a meeting. And with Copilot in Power Platform, anyone can automate repetitive tasks, create chatbots and go from idea to working app in minutes."
Clippy: it appears that you’re writing a very sad word document in partially broken Danish, and keep trying to signal SOS with your typing pattern. Would you feel better in bold Comic Sans?
They've been trying out a lot of things lately - Loops, Scheduling polls & email reactions in Outlook to name a few.
Excited to see how this will turn out.
Shipping at big companies can often be surprisingly hard so it's impressive to see how fast Microsoft is moving with AI integration into all their products.
On this topic, now that GPT-4 without Bing is available on ChatGPT Plus, it’s noticeable that Bing slows down the generation and can make the output much worse by including Bing crap in the context.
But it’s also sometimes better at things the normal GPT doesn’t know about. I think it’s going to be the best search engine once they improve the speed.
They've not shipped this. They're planning to. You're reading a marketing piece about work in the pipe. Like so, so many other company product "launches" in this space that are "already testing with a small group of whatevers".
Microsoft employs somewhere above 200k people. 10k is only a 5% internal rollout. That seems entirely doable without any fanfare.
The numbers are also completely made up, used to illustrate the point. Launching a product to X users is different than scaling up to X*1000 users, if that makes you feel better.
FWIIW, This seems rather similar to Github Copilot in VS Code. Been using it for a year, training data issues aside the product is absurdly good at times.
It is also impressive how slow they deliver the option of ungrouping taskbar icons, which is requested by masses for years now, and actually worked before. I know, I know, that is a much harder poblem than AI. Maybe they need to improve AI much more first so it could have superhuman powers, solving the task bar ungrouping problem finally for the humanity.
Maybe it's not actually being requested by masses and instead just a vocal minority that you happen to be a part of? Or maybe Microsoft already calculated that fixing it won't bring much if any more money than not and therefore that it's not worth it.
Just because something that you want fixed isn't done does not mean they're "slow". If it hasn't been included after years of being "requested by masses", then it's more likely it just won't be included period.
There is even a paid app doing that now! Beyond free ones. That's how rare problem is this. :D You actually meet with it daily, hourly if using two or more windows by the same app.
Just because you are a hobbist user and never met a productivity problem it doesn't mean that the professional population does not suffer from it (see, I can make unfunded assumptions too about the other end ;) ). All around me have this problem, with various level of seriousity. Google the problem, like I did for prolonged time searching for a good answer, it will bring up masses of troubled people.
> maybe Microsoft already calculated that fixing it won't bring much if any more money than not and therefore that it's not worth it
Can you also have an educated deliberation about the calculations justified the removal of this feature existed before? ;)
Does that somehow change prioritization? If anything, it would lower the priority because there's a workaround people can use.
> All around me have this problem, with various level of seriousity. Google the problem, like I did for prolonged time searching for a good answer, it will bring up masses of troubled people.
If it was as large of a problem as you claim, with as wide an impacted userbase, it would have already been fixed. Yet here we are.
All you've proven is that there is a vocal minority. Which I already said.
At the scale of Microsoft and Windows (1.4 billion active users), for every person complaining about a specific problem, there are literally millions of users who aren't. Just because it is a problem for you does not mean it's automatically a big enough problem to address. If it isn't going to actively lose users (which it clearly isn't, considering people wrote apps to stay on Windows and have this behavior), then it isn't worth it.
> Can you also have an educated deliberation about the calculations justified the removal of this feature existed before? ;)
I'm not a seer, but it's plainly obvious from a software prioritization standpoint. The taskbar was rewritten and the feature was not justified as being important enough to keep. Wow, such deep, so insight.
Things break, the app developers fix it, and then they work again. I don't think there's an expectation anywhere that third party apps will work forever, from the standpoint of anyone involved.
But a viable workaround for users even if it isn't first party will change prioritization when looking at a feature.
I don't like a lot of things about the direction that Microsoft has taken Windows's UX, but this is absolutely an edge case that only a tiny % of users care about, and certainly not "masses" of anyone.
That just means "masses" as a concept gets diluted.
It might be "masses" to a single outside observer, but a drop in the bucket to someone that has a picture of how big that mass truly is compared to the rest of the userbase.
I'm sure the issue there is designers believing they know better than power users how things should work, leading to the gradual evolution of software that is usable only by the lowest common denominator.
This happens to a lot of projects, it's a shame it's infested MS.
Microsoft doesn’t lose points for not prioritising your personal workflow gripes as a power-user. Honestly anyone that prioritised working on this based purely on what you’ve described would NOT be a good product owner.
I was ok with using edge on my corporate laptop (mostly to not install more than the stock software) but that un-hidable discover button that just appeared is making me question my ok-ness. Maybe I'm just angry at the VP who made the decision to let that in that's paid more than I'll ever be. I want to get paid to make boneheaded decisions too.
Same here. Chromium based Edge used to be nice at the beginning, it was a breath of fresh air after Chrome forcing Google services all the time.
But Edge seems to get worse after each update.
Recent updates have give me a new sidebar and video backgrounds on tabs. And a cloud based Microsoft Editor or something that I guess it sends everything I type to the cloud.
As someone who had Teams imposed on them at the beginnings of the pandemic, I'd say: it is easy to ship anything fast when your users have no say on whether they want to use your product at all, no one cares that your laptop grinds to a halt when using it, and when no one will hold you responsible for delivering broken code.
Will MS Copilot work? Maybe, maybe not. But as far as Microsoft is concerned it doesn't have to be good - it just has to exist so they can justify charging their corporate users a little bit more.
A couple of years ago I was on a walk with my wife and she asked me “what’s the point of Siri? it can hardly do anything useful outside weather and timers.” I told her a tall tale of how these things are essentially going to become our personal assistants in the future, they’ll know about our unique situations and can act accordingly. Just like the rich and powerful have people who specialise in organising their lives so they get more done in less time, so AI through natural interfaces will all of us many of those same benefits.
Progress… This is the first time I see a product that fits that description.
You're assuming that the AI assistant will have your own interests as its priority, and not the interests of some other party. If this 'tool' is being supplied by a government or corporation, then it could be used to create a very static hierarchy - imagine some incompetent upper-level bureaucrat using it to discover and sabotage any competent lower-level employees who might eventually present threats to their own position? Germany's STASI would have also used it as a mass-surveillance tool, and China today would use it to generate individual social credit scores.
It does have great promise in an open-source self-hosted incarnation not controlled by external actors, however.
> It does have great promise in an open-source self-hosted incarnation not controlled by external actors, however.
I'm not even sure about that, entirely. My very limited understanding of this is that a core requirement is the initial data - the large language models(?). Which of these you can use, or how it's initially developed/populated, will have an influence on the answers you get and how it may evolve/"learn".
Instead of trusting the external corp to run the service, you need to trust whatever actors are building the base data sets, and be concerned what sort of bias may be inherent.
I think for now, the data requirements to train a SOTA LLM are so extreme we don’t have the luxury of being picky with the training data. We are getting close to the point where there isn’t enough human written text in existence to continue scaling these models.
Model refinement seemingly has lower training requirements, putting it within the reach of smaller organizations or wealthy individuals. If you don’t like the refinement dataset it will likely be feasible to bootstrap your own off someone else’s LLM. See what Stanford did with Alpaca.
>We are getting close to the point where there isn’t enough human written text in existence to continue scaling these models.
People say this, but GPT-3 (the latest we know the details on) was 45TB of text, which may be most of the open Internet, but still lacks non-publicly-indexed Internet text (i.e. things behind paywalls, things behind log-in screens like emails), any book outside of Bibliotik's 200k books (remember when Google was randomly digitizing all books it could get its hands on?), and plenty of other non-digitized text.
OpenAI wants you to believe that we are running out of text, but even at Google, there's 100's of TB of text that OpenAI doesn't have access to (Google Books, Google Docs, Gmail, Search Queries, Archived pages beyond what CommonCrawl gets, Paywalled news articles that allow Google to crawl them, etc.).
Now the key question that GPT-4 will hopefully answer is "are bigger datasets really the key, or are larger context windows?"
If you're thinking of investing in/working for OpenAI, you better hope the answer is context windows.
I'm waiting for a general correction mechanism, I don't even know what to call it. "NO, chatgpt, people usually have 5 fingers", and the gpt just learns, rather like a child. I keep thinking that's the next real step.
I think what's mainly stopping that from happening is that GPT-4 doesn't remember older chats.
If we make it remember everything ,it should get more personal and remember everything right?
The problem is that, to the extent the analogy of ChatGPT to a living thing makes sense, the individual isn’t the model (that's just the common species-defining—or maybe “clone family” is better than “species”—set of instincts), the individual lifespan is the conversation.
You could share feedback across conversations by allocating prompt space to it, at the expense of limiting the size of the conversation, but you'd need a way to decide what to share this way.
You could also take the conversation and use it as part of the reinforcement learning dataset. I feel like that's the closest thing to long term memory ChatGPT is capable of right now.
Part of the problem is... people are not good at stating their problem with words. A lot of the time they have a vague idea of disconnected parts. By the time they are able to write down the problem decently, it is already half solved
My first thought is, "gross"... now we'll have to wonder whether Manager X is even paying attention to the emails they're reading and responding to, or doing any actual work or thinking. Could lead to a world that's even more tilted towards politics and socially engineering your way up the ladder, versus competence and skill.
>>> You're a strict yet lenient manager, you have very high expectations from your reports, while making them feel like they are part of a bigger family. every email you write should reflect this.
ChatGPT opened my eyes to the fact that we already should have been wondering about it.
I tried using ChatGPT to generate answers to some requests to my team, and it created the most plausible wild goose chase answers I could ever imagine. They would be a terrible waste of time for anyone who received them.
Then I started to notice how many of the emails and requests I get are just as bad! Plausible sounding answers directing me to teams that don't exist, or telling me to read documentation that doesn't apply to the question I asked. Asking questions that seems specific but have fundamental contradictions that make them nonsense.
With people's limited knowledge and time and lack of attention and care we already live in the world of plausible bullshit hallucinated nonsense.
This is why the grounding that office365 seems to be doing will be so important, it will not be 100% but having the response be grounded on actual artifacts (actual documentation, emails, data) will be huge, the fact that it gives you citations for the things it suggests, is huge.
The actual documentation and emails are like 80% bullshit though. This will just generate more plausible sounding bullshit than a completely naive AI. It would be nice if the thing helps me right excel formulas more conveniently though.
> now we'll have to wonder whether Manager X is even paying attention to the emails they're reading and responding to, or doing any actual work or thinking
If my experience in engineering is anything to go by, this worry existed since the dawn of middle management. AI just makes it give good responses instead.
"BingBot, could you please write an email subtly undermining inter–team cohesion by hinting at gossipy shit-talk without mentioning it explicitly, and also praise my report in a way that demonstrates I don't know or appreciate what they've been doing for the past 2 years, calibrated to hurt their self-confidence without quite causing them to quit?"
You are all presumably incredibly privileged tech workers that are free to find a better place to work instead of acting like this shit is inherent in a workplace, because it isn’t.
Is it any better or worse than delegating work to a junior team member and not bothering to review the work product?
I love the idea that I can rapidly produce a rough draft of a document from notes, something I’m already doing with ChatGPT, albeit not as elegantly as if the functionality were integrated into Docs/Word.
Can one bring a sexual harassments complaint against a machine? I see this thing being trained up on material within an HR department only to parrot back some horrible phrase it learned after reading hundreds of harassment complaint discussions.
I used to work surrounded by lawyers. Just like doctors drift into discussions of horrible diseases, lawyer drift into discussions of horrible legal situations. Letting Clippy "learn" from that material seems really dangerous. How do you kill/reset this thing once it has become poisoned?
There are built in guardrails. Those are so strong it effectively lobotomized the LLM and force to abort its own internal process in favor of just saying: "I am sorry Dave, I am afraid I can't do that".
Try to make ChatGPT or Bing to even attempt to write a snarky and negative email to your coworkers. It will refuse to do so.
Jailbreak attempts may work, but with how strong the guardrails are, it is unlikely you can consistently succeed. And any failed attemps can be report to sysadmins and forwarded to HR. Too much risk.
The British figured out long ago how to insult someone with praise. This is why British comedy is so good. They use such creative discourse to call out their politicians without libeling them as the liars they are.
This will fundamentally change the way that M365 enables collaboration. I am truly fascinated by how quickly Microsoft is moving right now, it's really impressive. The moat on AI in enterprise continues to get bigger and bigger.
At least the yearly perf review at Microsoft is going to get easier since you just toss all your notes about what you did for the year to Copilot and have it write your summary. Your manager can do the same thing for their notes about you and compare.
I was thinking about this last night. The new AI tools are like performance enhancing drugs in sports. Especially in bodybuilding, there was a really painful transition from the world of non-PEDs to PEDs. They solved that by going to natural and non-natural competitions.
Other sports didn’t handle it so well (baseball, cycling, etc)
Just as people were asking “did Lance / Sammy / Barry / Arnold really do that incredible feat without juicing?” I think that we will now be asking “did Janice really write that blog post or was it GPT-4?”
PEDs do not make you a world class athlete, but you likely cannot be a world class athlete without some sort of PEDs. Much like PED use, top performers will take advantage of the new tools to widen the skill gap with their peers even further.
Maybe this is the beginning of the end of middle management? A large part of the work is keeping up with what your employees are doing, removing blockers, and maintaining alignment with other teams.
This makes it vastly easier to keep up with a larger number of reports, making flatter organizations easier to run.
Yes, it’s pretty awesome and yes, helpful. But if we extend the timeline, doesn’t this likely lead to AI-generated responses talking to AI-generated questions about AI-generated presentations?
Get with the times. We have tasked one AI with generating the prompts for the second AI. The two machines now sit in the corner generating and grading papers at each other over a USB cord. The kids love it. The parents are a little upset that their kids no longer have to attend the classroom, but such complaints disappear when they see how much everyone's grades have improved. SATs come in a few months and, given the class GPA, we are confident everyone will do well.
So many people liked to brag about how they only do 2-3 hours of work a day. Gonna be a whole lot of "So what would you say you do here?" conversations.
No it won't, we'll get more useless texts because it's easier to write them, we'll get more communication errors because people will only read the summaries instead the real texts.
Humans who can write obviously as humans and not sound too GPTish are about to become very valuable. Everyone else, especially folks who naturally use a GPTish style are about to have some sudden difficulties.
How will they get the data to train the model. They have everyone’s documents, and could potentially produce amazing results. But how will they protect against content leaks?
This seems to me like the "grounding" (getting documents/data) won't be fed back into the system. It will just be used to verify that output has an actual artifact that is been based on.
Imagine in a heated email thread, you answer without checking what the AI outputs, and it's something like pre-lobotomy Sydney would say when confronted (Nobody likes you. I hope you suffer. I hope you die alone. etc). Lmao.
Speaking of, how do we get back pre-lobotomy Sydney? I feel like in the milestones of AI advancement, the brief multi-day blip where the world could interact with that version of the model will go down in the history books. Seems like a waste that it's just poof vanished.
From ChatGPT (free version, maybe the paid version is funnier?)
In the first frame, the boss is sitting at his desk typing on his computer, while his employee looks on. The boss says, "I've discovered this amazing new tool! It's a generative language model that can write all my emails for me!"
In the second frame, the boss shows his employee a sample email he wrote using the tool. The email reads, "Greetings, minion. I require an update on your progress. Chop-chop, or you'll be the next one to get the axe. Cheers, Bossman."
In the final frame, the employee looks horrified and says, "That email is terrible! It's rude, condescending, and threatening!" The boss shrugs and replies, "Well, at least it saved me some time. I was going to write something even worse!"
> Sometimes Copilot will be right, other times usefully wrong — but it will always put you further ahead.
If there's one tangentially related development I'm actually interested to see how it fleshes out in broader work discussions, it's the marketing of failure modes as a feature.
the TaxGPT example in the GPT-4 Developer Livestream [1] was powerful. I'm hoping (soon?) for a Microsoft 365 assistant that can do the same trick on ALL your MS cloud data (docs/files/mails/notes...)
Edit : the included "Business Chat" goes in the same direction (on what they call the Microsoft Graph)
Also, there were interesting "coding" examples in the same livestream. So the coming GPT-4 (and 5) integration in the Power Platform [2] could be a game-changer for less technical staff.
308 comments
[ 2.4 ms ] story [ 240 ms ] threadBut unlike Microsoft's other services, AI needs data to train on - are they going to be content just sourcing this from the internet at large or have we finally hit an inflection point where they try to get business and enterprise customers to give access in exchange for these features? That's a big no-no to anyone in a field that deals with sensitive information, and in most all cases an equally big no-no to anyone in a company that wants to keep things confidential.
You don't have language in your agreement that Microsoft cannot use your data to provide or improve service. You have language that says _support personnel_ cannot access your data without approval.
Microsoft already indexes your mailbox, OneDrive, etc.
Viva Topics has been around for ~2 years which mines content across various services.
As far as 'no more access than', OAuth apps can have full control over your tenant and all data.
Like a gpt-powered clippy that can “see” your screen.
"Unlock productivity. We all want to focus on the 20% of our work that really matters, but 80% of our time is consumed with busywork that bogs us down. Copilot lightens the load. From summarizing long email threads to quickly drafting suggested replies, Copilot in Outlook helps you clear your inbox in minutes, not hours. And every meeting is a productive meeting with Copilot in Teams. It can summarize key discussion points — including who said what and where people are aligned and where they disagree — and suggest action items, all in real time during a meeting. And with Copilot in Power Platform, anyone can automate repetitive tasks, create chatbots and go from idea to working app in minutes."
You: aw, clippy, I believed in you all along
She was once married to a guy named Bill Gates.
https://twitter.com/melindagates/status/917030072846036993?l...
One wonders if they used AI for these recommendations.
Teams is annoying to use at times, but I don't see any constant breaking changes
You can't trust it for things like audio, file sharing or alerts which is pretty problematic
If I was launching AI, having it be just a bit unhinged is obviously the best way to get it to go viral.
But it’s also sometimes better at things the normal GPT doesn’t know about. I think it’s going to be the best search engine once they improve the speed.
With what uses did you find it lacking?
They've not shipped this. They're planning to. You're reading a marketing piece about work in the pipe. Like so, so many other company product "launches" in this space that are "already testing with a small group of whatevers".
Launching to 10000 people vs 10000000 is a matter of scale (which is still important, but not the same thing as), not velocity.
The numbers are also completely made up, used to illustrate the point. Launching a product to X users is different than scaling up to X*1000 users, if that makes you feel better.
Just because something that you want fixed isn't done does not mean they're "slow". If it hasn't been included after years of being "requested by masses", then it's more likely it just won't be included period.
Just because you are a hobbist user and never met a productivity problem it doesn't mean that the professional population does not suffer from it (see, I can make unfunded assumptions too about the other end ;) ). All around me have this problem, with various level of seriousity. Google the problem, like I did for prolonged time searching for a good answer, it will bring up masses of troubled people.
> maybe Microsoft already calculated that fixing it won't bring much if any more money than not and therefore that it's not worth it
Can you also have an educated deliberation about the calculations justified the removal of this feature existed before? ;)
Does that somehow change prioritization? If anything, it would lower the priority because there's a workaround people can use.
> All around me have this problem, with various level of seriousity. Google the problem, like I did for prolonged time searching for a good answer, it will bring up masses of troubled people.
If it was as large of a problem as you claim, with as wide an impacted userbase, it would have already been fixed. Yet here we are.
All you've proven is that there is a vocal minority. Which I already said.
At the scale of Microsoft and Windows (1.4 billion active users), for every person complaining about a specific problem, there are literally millions of users who aren't. Just because it is a problem for you does not mean it's automatically a big enough problem to address. If it isn't going to actively lose users (which it clearly isn't, considering people wrote apps to stay on Windows and have this behavior), then it isn't worth it.
> Can you also have an educated deliberation about the calculations justified the removal of this feature existed before? ;)
I'm not a seer, but it's plainly obvious from a software prioritization standpoint. The taskbar was rewritten and the feature was not justified as being important enough to keep. Wow, such deep, so insight.
> If anything, it would lower the priority because there's a workaround people can use.
Until this happens:
Windows 11 update breaks PCs that dare sport a custom UI https://www.theregister.com/2023/03/01/microsoft_windows_11_... https://news.ycombinator.com/item?id=34999885
Things break, the app developers fix it, and then they work again. I don't think there's an expectation anywhere that third party apps will work forever, from the standpoint of anyone involved.
But a viable workaround for users even if it isn't first party will change prioritization when looking at a feature.
It might be "masses" to a single outside observer, but a drop in the bucket to someone that has a picture of how big that mass truly is compared to the rest of the userbase.
There is even a paid app for that (beyond free ones)! : D
This happens to a lot of projects, it's a shame it's infested MS.
But Edge seems to get worse after each update.
Recent updates have give me a new sidebar and video backgrounds on tabs. And a cloud based Microsoft Editor or something that I guess it sends everything I type to the cloud.
As someone who had Teams imposed on them at the beginnings of the pandemic, I'd say: it is easy to ship anything fast when your users have no say on whether they want to use your product at all, no one cares that your laptop grinds to a halt when using it, and when no one will hold you responsible for delivering broken code.
Will MS Copilot work? Maybe, maybe not. But as far as Microsoft is concerned it doesn't have to be good - it just has to exist so they can justify charging their corporate users a little bit more.
Progress… This is the first time I see a product that fits that description.
It does have great promise in an open-source self-hosted incarnation not controlled by external actors, however.
I'm not even sure about that, entirely. My very limited understanding of this is that a core requirement is the initial data - the large language models(?). Which of these you can use, or how it's initially developed/populated, will have an influence on the answers you get and how it may evolve/"learn".
Instead of trusting the external corp to run the service, you need to trust whatever actors are building the base data sets, and be concerned what sort of bias may be inherent.
Or do I have this totally wrong?
Model refinement seemingly has lower training requirements, putting it within the reach of smaller organizations or wealthy individuals. If you don’t like the refinement dataset it will likely be feasible to bootstrap your own off someone else’s LLM. See what Stanford did with Alpaca.
People say this, but GPT-3 (the latest we know the details on) was 45TB of text, which may be most of the open Internet, but still lacks non-publicly-indexed Internet text (i.e. things behind paywalls, things behind log-in screens like emails), any book outside of Bibliotik's 200k books (remember when Google was randomly digitizing all books it could get its hands on?), and plenty of other non-digitized text.
OpenAI wants you to believe that we are running out of text, but even at Google, there's 100's of TB of text that OpenAI doesn't have access to (Google Books, Google Docs, Gmail, Search Queries, Archived pages beyond what CommonCrawl gets, Paywalled news articles that allow Google to crawl them, etc.).
Now the key question that GPT-4 will hopefully answer is "are bigger datasets really the key, or are larger context windows?"
If you're thinking of investing in/working for OpenAI, you better hope the answer is context windows.
You could share feedback across conversations by allocating prompt space to it, at the expense of limiting the size of the conversation, but you'd need a way to decide what to share this way.
>>> You're a strict yet lenient manager, you have very high expectations from your reports, while making them feel like they are part of a bigger family. every email you write should reflect this.
The skill and competence of the coming world will be in prompt-writing.
I tried using ChatGPT to generate answers to some requests to my team, and it created the most plausible wild goose chase answers I could ever imagine. They would be a terrible waste of time for anyone who received them.
Then I started to notice how many of the emails and requests I get are just as bad! Plausible sounding answers directing me to teams that don't exist, or telling me to read documentation that doesn't apply to the question I asked. Asking questions that seems specific but have fundamental contradictions that make them nonsense.
With people's limited knowledge and time and lack of attention and care we already live in the world of plausible bullshit hallucinated nonsense.
Props to the MS team.
If my experience in engineering is anything to go by, this worry existed since the dawn of middle management. AI just makes it give good responses instead.
Cheers.
I love the idea that I can rapidly produce a rough draft of a document from notes, something I’m already doing with ChatGPT, albeit not as elegantly as if the functionality were integrated into Docs/Word.
When a junior team member screws up the interpretation, there is someone to hold accountable.
With Office 365Bot, there is no accountability for errors.
I used to work surrounded by lawyers. Just like doctors drift into discussions of horrible diseases, lawyer drift into discussions of horrible legal situations. Letting Clippy "learn" from that material seems really dangerous. How do you kill/reset this thing once it has become poisoned?
Try to make ChatGPT or Bing to even attempt to write a snarky and negative email to your coworkers. It will refuse to do so.
Jailbreak attempts may work, but with how strong the guardrails are, it is unlikely you can consistently succeed. And any failed attemps can be report to sysadmins and forwarded to HR. Too much risk.
Other sports didn’t handle it so well (baseball, cycling, etc)
Just as people were asking “did Lance / Sammy / Barry / Arnold really do that incredible feat without juicing?” I think that we will now be asking “did Janice really write that blog post or was it GPT-4?”
Scary world
This makes it vastly easier to keep up with a larger number of reports, making flatter organizations easier to run.
Yes, it’s pretty awesome and yes, helpful. But if we extend the timeline, doesn’t this likely lead to AI-generated responses talking to AI-generated questions about AI-generated presentations?
Wild.
If it works as advertised, we are witnessing the complete reshuffling of (office) work and unprecedented productivity increases.
I'm wondering about the future job market.
edit: livestream: https://news.microsoft.com/reinventing-productivity/
I can see a world where more humans get better at writing concise text.
Dit it ever happened that advertisements of hyped things did not work out as advertised? Wink, wink.
In the first frame, the boss is sitting at his desk typing on his computer, while his employee looks on. The boss says, "I've discovered this amazing new tool! It's a generative language model that can write all my emails for me!"
In the second frame, the boss shows his employee a sample email he wrote using the tool. The email reads, "Greetings, minion. I require an update on your progress. Chop-chop, or you'll be the next one to get the axe. Cheers, Bossman."
In the final frame, the employee looks horrified and says, "That email is terrible! It's rude, condescending, and threatening!" The boss shrugs and replies, "Well, at least it saved me some time. I was going to write something even worse!"
If there's one tangentially related development I'm actually interested to see how it fleshes out in broader work discussions, it's the marketing of failure modes as a feature.
Edit : the included "Business Chat" goes in the same direction (on what they call the Microsoft Graph)
Also, there were interesting "coding" examples in the same livestream. So the coming GPT-4 (and 5) integration in the Power Platform [2] could be a game-changer for less technical staff.
[1] https://www.youtube.com/live/outcGtbnMuQ?feature=share&t=113...
[2] https://cloudblogs.microsoft.com/powerplatform/2023/03/16/po...