While the content of AI generated images/videos will be hard to distinguish from its real one's, the relationship of content to its metadata will be hard to fake, in particular with current LLMs, where this relationship is not used in the training process.

Therefore, I believe MS will make quite a bit of loss per user and needs to get a lot of licenses to ppl who actually don’t use Copilot a lot or it can significantly lower the costs for running the models (unlikely in the short term).

Azure OpenAI GPT-4 costs for customers is $0.06 per 1000 input tokens (ignoring output tokens for now), if we assume a 30% margin for MS (very unlikely) that means $0.046 cloud costs for MS, i.e., that Copilot user is already causing MS ~$45 costs by only using Copilot in Teams.

Let’s assume ppl speak 130 words a minute, i.e., 7800 words an hour, translates to about ~11000 tokens in English. Let’s assume 4h of meetings for a manager a day, 22 working days a month and only 1 Copilot call per meeting (very unlikely) which translates to 968000 input tokens.

Taking the example of a manager, they will love Copilot for Teams to get summarisations and extractions of their meetings. Managers will have a lot of meetings, so the costs for MS to serve Copilot are as follows:.

MS is charging $30 per user, with a min. of 300 users per company, i.e., companies will first make Copilot available to staff who are potential heavy users, e.g., marketers, HR, customer service agents, managers etc.

Microsoft has released Copilot today and the stock market has been very excited about the potential revenue increase for MS. However, what I think is rather overlooked is the costs for running Copilot per user for MS.

GA for MS 365 #Copilot is on the 1st of November. I have early access already and while there is some really cool stuff (Teams call summarisation), I have created the following simple flowchart to help one decide if to use Copilot or if a custom GenAI solution serves you better.

I expected Apple to come out with a big GPU change to their #iPhone15 to run Foundational Models on them. Does not lool like they have. Will need to try, but it appears as if they missed a massive opportunity.

The collapse of SVB and CS poses the question again why do we need banks as deposit holders and creditors at all? Why are we not depositing/loaning money at/from central banks directly. Nowadays without branches, the services banks offer may as well be offered by central banks.

This is why I am such a proponent of open reviews in CS, as is successfully done in the @SW_journal.

Happy to announce that our article on Active Knowledge Graph Completion has finally been published in Information Science. Thanks to Pouya, who persisted in convincing reviewers that learning questions is at least as important as learning answers in KGs!.

Our paper on 'Links in Wikidata', accepted at @eswc_conf 2022, is now available as pre-print at: @AxelPolleres @sjrm_142857.

RT @hackingbutlegal: This, out of #Ukraine, is 100% one of the most incredible videos I have ever seen. This Russian POW has the heart of….

RT @VeraMBergen: I can’t wrap my mind around the existence of this clip.

The morale of Russian troops must approach 0. It has to get negative for them to return to Moscow and kick Putin out of his palace!.

Putin, remember, if you wipe out humanity, there is no history to remember you.

So Europe is still taking Russian gas. Sanctions that hurt Russia would inevitably hurt also Western countries. But they are not willing to do so. So much to #standwithukraine