Sign in

Alex

@alexo.bsky.social
130 followers 335 following 64 posts

AI-skeptic AI researcher. Director of the Centre for Analytics & AI Engineering at the University of Toronto

PostsRepliesMedia
Reposted by Alex
Matt Elliott @graphicmatt.com · 06/10/2026
Alexander dropping out will have people asking whether this benefits Bradford. Luckily, @alexo.bsky.social covered this very question in this week’s edition City Hall Watcher. The short version: it probably doesn’t mean that much. toronto.cityhallwatcher.com/p/where-toro...
Stacked bar chart from Mainstreet Research (September 28-29, 2026) showing polling results for a two-candidate choice between Chow and Bradford, with Chow leading at 47.1% versus Bradford's 40.9% in the head-to-head matchup.
2267
Reposted by Alex
Matt Elliott @graphicmatt.com · 05/10/2026
Check your inbox: it’s City Hall Watcher #404! Three weeks out, where does the mayoral election stand — and how might it change? @alexo.bsky.social looks at the state of the race. Plus: LOBBYIST WATCH looks at garbage trains and farming plans. toronto.cityhallwatcher.com/p/where-toro...
2102
Alex @alexo.bsky.social · 25/09/2026
It’s a good question, and polls are inconsistent about “other”, “undecided” etc. The model factors this in, and looks at past elections + their polling to consider what might happen to this share. That is (part of) the reason we see Bradford winning around 1 times in 4
010
Alex @alexo.bsky.social · 25/09/2026
Gong is in this one!
100
Alex @alexo.bsky.social · 23/09/2026
That’s how we are handling it in the actual model - it’s still interesting because it’s quite out of line with the other polls around the same time. Different survey methodology & different question phrasing at work? But all polls point to some kind of Chow lead.
010
Alex @alexo.bsky.social · 23/09/2026
Aside from the Ipsos poll which was just done somewhat differently; the polls are all consistently(-ish) showing about a ten point lead for Chow. When we see that “narrow” in a couple polls, it looks more like random variation in the sample.
020
Alex @alexo.bsky.social · 07/09/2026
School of cities maybe?
011
Alex @alexo.bsky.social · 29/08/2026
There’s no road tax any more… cyclists are paying for road infrastructure just as much as anyone else
170
Alex @alexo.bsky.social · 28/08/2026
Yeah, this is reasonably accurate. Remarkable display of the gap between capability and judgement
000
Alex @alexo.bsky.social · 28/08/2026
The John Cleese who uses his platform to vocally back far-right British politicians? www.telegraph.co.uk/news/2026/07...
telegraph.co.uk
110
Alex @alexo.bsky.social · 30/07/2026
BlogTO published this with the headline "Ferry rides to the Toronto Islands will be free in August but only for select people". Isn't that remarkable
010
Alex @alexo.bsky.social · 26/07/2026
If I’m going to defend this movie at all, I appreciated Spielberg’s ability to construct and film a really complicated sequence! But I still agree with you it was a bad movie
110
Alex @alexo.bsky.social · 26/07/2026
I’m serious but open to disagreement! Trying to find something positive in a flop
100
Alex @alexo.bsky.social · 26/07/2026
The train sequence was a silver lining, surely?
100
Reposted by Alex
Technologists for Democracy @techfordemocracy.bsky.social · 20/07/2026
Technologists for Democracy is calling on the Minister of Transportation in Ontario (MTO) to reject Waymo's planned entry into our province. Read our open letter to the MTO for more details. Sign the petition to Say No to Waymo techfordemocracy.ca/2026/04/29/s...
techfordemocracy.ca
Tell the MTO: Say NO to Waymo
Waymo is coming to Canada, and we believe they will cause issues around digital sovereignty, job loss, and accountability.
199
Alex @alexo.bsky.social · 17/07/2026
I have to admit I'm not a regular at Islington station, and I know it's under construction, but following all the posted signs for Islington Avenue (west side) led me to an impassable bank of bins and a metal fence. What are we doing here?
Signs direct those looking for Islington Avenue West down a temporary construction path……but that path leads to a dead end obstructed by bins.
040
Reposted by Alex
Rachel Gilmore @rachelgilmore.bsky.social · 15/07/2026
Got upset and made this so hopefully people will actually listen. What the fuck are we doing here folks?
881460793
Reposted by Alex
Matt Elliott @graphicmatt.com · 25/06/2026
Quick video of Councillor Rachel Chernos Lin's driving-while-talking-about-road-safety moment at Council.
1816936
Reposted by Alex
The Biking Lawyer LLP @thebikinglawyer.bsky.social · 24/06/2026
Same day, different white pickup, near rush hour, blocking the bike lane and sidewalk. Perhaps there was a real emergency here that required this move and I hope he's ok, but there did not appear to be a crash and dude was calmly on his phone. Do we live in a society?
4143
Alex @alexo.bsky.social · 11/06/2026
AWS is also used for a lot more than just AI
080
Alex @alexo.bsky.social · 05/06/2026
Community Safety HUB
010
Alex @alexo.bsky.social · 28/05/2026
The transparency was part of the work from the offset! The human-backed validation was something we developed along the way, but in line with scientific methodology for this type of work
010
Alex @alexo.bsky.social · 27/05/2026
Happy to go into more detail or elaborate if you have follow ups
100
Alex @alexo.bsky.social · 27/05/2026
The other part is that LLMs write one word at a time, with no way to go back. And each word it “picks” has a certain amount of randomness to it. So the model can, by bad luck, almost “blurt out” that Google has two Ps, and at that point it’s stuck with that answer
110
Alex @alexo.bsky.social · 27/05/2026
Couple of things - one, as often discussed, is that the model doesn’t see “Google” as a word made of letters - it sees it more like a set of co-ordinates pointing to the underlying meaning it’s learned (so probably close to “company” in one direction, “tech” in another)…
110
Alex @alexo.bsky.social · 01/05/2026
Great quote from neuroscientist Uri Hasson: “My worry is not that these models are similar to us. It’s that we are similar to these models.”
1131
Alex @alexo.bsky.social · 27/04/2026
Yeah, it’s super interesting the different fields that have been tackling this question! I have some notes but I haven’t had time to write them up yet.
110
Alex @alexo.bsky.social · 22/04/2026
It would be great if this tracked what I’ve already seen vs what’s new - if you open-source it I could work on that
020
Alex @alexo.bsky.social · 21/04/2026
There’s no doubt that we have the cart before the horse in terms of showing that these models can perform without fully being able to explain why. I think it’s the fact I agree with you on this that makes me sensitive to the areas where we actually do have explanations.
100
Alex @alexo.bsky.social · 21/04/2026
It’s the principle of doing the scientific research before making the claim. Trust me, I am frustrated by how anti-scientific much of the “research” coming from the AI industry is (Anthropic “publishes” “papers” that have not undergone any form of peer review) - let’s be better.
120
Alex @alexo.bsky.social · 21/04/2026
Happy to do it - I think it’s a really interesting question and worth a review either way. If I provide you with those attempts, will you retract your claim that nobody has an explanation?
210
Alex @alexo.bsky.social · 21/04/2026
Would you like me to do a short literature review to evaluate the merits of the claim you made?
100
Alex @alexo.bsky.social · 21/04/2026
(I’m being facetious - I don’t think you did because a cursory google search found that in fact, someone has an explanation)
100
Alex @alexo.bsky.social · 21/04/2026
I just thought your significant claim that “nobody has any explanation” might warrant a review of the literature. Did you happen to do one?
100
Alex @alexo.bsky.social · 21/04/2026
Your personal ignorance of the science actually undermines the point you are looking to make. Real research explains the phenomenon, but also shows it harms accuracy: arxiv.org/abs/2603.18507
arxiv.org
Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM
Persona prompting can steer LLM generation towards a domain-specific tone and pattern. This behavior enables use cases in multi-agent systems where diverse interactions are crucial and human-centered ...
110
Alex @alexo.bsky.social · 11/04/2026
Apps like tinder will actually penalize you for swiping right as fast as you can on everyone. But that doesn’t mean they’re giving anyone a fair shake - they effectively decide how many matches you’ll get in a given period before you even interact. It’s all about making you pay for premium.
041
Alex @alexo.bsky.social · 16/03/2026
Yikes, ventilation must be really bad at the revue. Are those CO2 levels safe? A mask will protect you from disease but not CO2
002
Alex @alexo.bsky.social · 17/02/2026
Later in the article she says she felt safer getting her children an unlicensed vaccine from “BabyJabs” than the licensed vaccine from the NHS
020
Reposted by Alex
David Zipper @davidzipper.bsky.social · 17/12/2025
New research finds that car touchscreens are so distracting that they induce "significant performance degradation" among drivers. Meanwhile, in-vehicle touchscreens remain completely unregulated in the US. doi.org/10.1145/3746...
28526164
Alex @alexo.bsky.social · 13/11/2025
If I’m reading this correctly, *directly next to Dufferin mall* is considered a “neighbourhood corner store”?
020
Alex @alexo.bsky.social · 13/10/2025
Do the existing e-bikes support it (maybe with modification)? Or would they need to get entirely new bikes
000
Alex @alexo.bsky.social · 19/06/2025
Now we just have to hope it never gets hot in May!
010
Alex @alexo.bsky.social · 08/05/2025
I asked @juliedzerowicz.bsky.social before the election whether the feds approve U.S. military participation in the Toronto Airshow. Her office just replied saying they can’t answer because Cabinet hasn’t been named, and to ask again “in a few months” (after the airshow?)
000
Alex @alexo.bsky.social · 05/05/2025
They have been doing that for a long time. Early versions of GPT used Reddit comments sections to judge the quality of a response during model training.
010
Alex @alexo.bsky.social · 04/05/2025
Although this is recommended, it’s actually not allowed for a car to run over a pedestrian because they aren’t pointing.
100
Alex @alexo.bsky.social · 04/05/2025
Cars are significantly heavier than bicycles, and bring several orders of magnitude more energy into a collision even at the same speed as a bike.
110
Alex @alexo.bsky.social · 29/04/2025
Provinces can’t unilaterally secede - see here: www.canlii.org/en/ca/scc/do... “secession of a province "under the Constitution" could not be achieved unilaterally, that is, without principled negotiation with other participants in Confederation within the existing constitutional framework.”
canlii.org
040
Alex @alexo.bsky.social · 29/04/2025
The "leading" numbers really don't mean much. This is evidence of that
000
Alex @alexo.bsky.social · 28/04/2025
You’re absolutely right about how image “generation” works but it is a little different for text since it’s word by word. The quoted post is right, though, LLMs are trained to predict sequences of words that seem likely - nothing to do with truthfulness
010
Alex @alexo.bsky.social · 23/04/2025
It’s like delivery bike riders are completely invisible. Saying “it’s all delivery e-bike riders” as if that means it’s not worth it - okay, is everyone going to stop using DoorDash too?
120