What’s your AI prediction for 2027?
Make it specific enough to be wrong!
One thoughtful comment unlocks the ability to create threads.
22 comments
Do you think that the replacement will be absolute or will routers start playing a much bigger role? I do agree with the outlook though - even right now for most tasks frontier models are overkill
Another prediction: nobody will select their own model by the end of 2027, it feels pretty archaic to do so even today!
Even with the routers though I think white collar work will use the highest thinking level ~0% of the time.
Agree, and I think the popularity of high speed inference providers (like groq, cerebras etc) and diffusion models will grow as users chase token speed.
I think we could see continued supply chain issues, datacenter moratoriums, and funding drying up for AI investment leading to worsening slowdowns. Despite that, I think we may see advances in robotics as @Kenny mentioned coupled with better application of existing AI which will make it feel like things didn’t slow down as much as they did. Assuming funding doesn’t dry up, I think we may also see some really stupid and excessive laws passed in the US, despite potentially good intentions. I hope we see legislative action, but I really hope I’m wrong about them being stupid and excessive. Lastly, if the global economy continues trending downward as seems to be right now, I think we may see AI-assisted distribution of recovery resources (funding, debt relief, credits, etc.). Not sure if it will be disclosed as such or not and I hope it’s not an old guy prompting Grok, but I digress.
In the C-Suite, I believe in 2027 AI will significantly streamline decision-making. There are already systems that do a decent job of scheduling 1:1’s and accepting meetings, but complex meeting scheduling (and the willingness to allow it) is still out of reach. On top of that, meetings will only be scheduled as they are needed, using AI judgement to determine if it aligns with priorities or should be dealt with at a lower level. The decisions that executives need to make will come with a clear issue and 3-4 potential adjustable mitigations with confidence levels and implications. This will free up executive attention to drive the ideas that we hired them to come up with instead of sitting in yet another review.
Meeting scheduling for 3+ people has been annoying hard for us to automate with AI, especially for meetings that include external participants like panel interviews. Looking forward to no longer doing this by hand!
Closest I’ve gotten is determining availability using the outlook (yuck) connector in ChatGPT. Gets us half the way there, but there’s still tons of opportunities to not just speed up meeting scheduling, but to reduce the number of meetings scheduled.
We will see the release of one or more open models (likely from China) that are capable of finding & exploiting catastrophic security vulnerabilities. This may only come after obliteration of a powerful model, by the community.
This will lead to a significant increase of attacks from bad actors and will require increasingly powerful closed models. This will be a significant driver of revenue for the frontier labs.
Yeah I really don’t see any solution for this without extreme centralization given that people can easily remove safeguards from open models.
Even with frontier models ~6 months ahead, that’s nowhere near long enough for everyone to upgrade their systems.
I think long term this might look like everyone being forced onto a handful of cloud providers who can provide adequate security (somewhat analogous to how only Cloudflare and a few others are capable of providing adequate DDOS protection to highly targeted sites).
a fun one: some companies will start paying people to not use ai at all. human involvement becomes a premium feature
Everyone is looking/waiting for RSI, nothing new there. Robots? My girlfriend wants one just to clean and cook, lol. I am all about security, so for me, its a way to protect everything, big Corps backends, an AI waiting in the shadows to jump on an attacker. Anthropic is slowly getting there, I received an email last night and filled out their forms, and am bound by the agreement that I know I cant say anything else and have probably divulged more than I already have.. There are a LOT of bad actors out there, I am worried, like a lot of people are about infras going down and I think AI can help with this, but are we late to the punch?
So, RSI, SECURITY AI, Robots are my top 3, sir.
I have a feeling that productivity industry converging on a “work app” idea - always on agents, documents, meeting notataker, dictation and tool connectors. Every major productivity app will look the same - hybrid of codex + notion + granola + wispr.
Frontier models will no longer be used for normal white collar work by the end of 2027, most work will be done by much cheaper faster models.
Frontier models will be reserved for generating new knowledge (e.g. science / math) and in adversarial domains (e.g. cybersecurity / trading).
The vast majority of work is already done by models slightly off the frontier (Opus gets much more usage than Fable), but I think by the end of 2027 most office work will be done by the equivalent of Haiku.