omermano.bsky.social
Data Scientist, former neuroscientist, Jersey City resident
32 posts
44 followers
524 following
Active Commenter
comment in response to
post
The massive investments into frontier AI only make financial sense if you think we'll hit recursive self-improvement soon and you think getting there first will let you take over the world (or at least be part of the negotiations of how to split it up).
comment in response to
post
Haaretz English or Hebrew edition?
comment in response to
post
Yup. I totally agree and I suspect there is an incentive problem here. The benefits of modernization are diffuse and difficult for politicians to sell. If the MTA solved this problem in a non-technical way, it might make it more difficult to get funding for modernization.
comment in response to
post
Modernizing signaling has benefits that go well beyond this particular merge. It should be a cheap way to reduce ongoing costs and improve service. The issue is with the current contracting process/environment we're not able to do it cheaply or fully take advantage of the benefits.
comment in response to
post
You're an expert and I'm not!
comment in response to
post
Discussing complicated topics on Bluesky is hard, sorry... I would like to push politicians to prioritize efficiency so we can get a better transit system for the cost. This project seems like an example. It seems like other agencies are able to do better. Do you think otherwise?
comment in response to
post
Did some more digging and it looks like there is a big range of costs for modern projects (as you would expect) $100M/km is in the high end. For instance line 3 is $45M/km. It wasn't fair of me to use their most efficient numbers from 30 years ago, but I still think we are inefficient by comparison.
comment in response to
post
If I'm reading the summary right, there is an additional $230m in fleet costs. The inflation adjusted cost of the Madrid 1995-1999 Program was about $65m per mile :https://worksinprogress.co/issue/how-madrid-built-its-metro-cheaply/
comment in response to
post
In Madrid that would get you about 10 miles of a new track with a new station every mile.
comment in response to
post
The google AI summary feature needs to include a "sanity check" step. LLMs have all the knowledge necessary to provide a correct answer if you don't confuse them with wrong information from the web. Here is an example from Gemini 2.0 advanced (which doesn't have search capability):
comment in response to
post
The ZipRecruiter error is probably due to an issue with a non-ML-based parser. Here is an example adjunct posting: archive.ph/lwBM1. The stipend is for a semester but the parser thinks it's per month. An LLM-based parser would probably be able to avoid the error. Gemini gets this right for instance
comment in response to
post
It sounds like you're looking for the successor to the Alphasmart Dana. Phones do have pretty good battery life if you turn off the mobile network. www.pencomputing.com/palm/Pen47/d...
comment in response to
post
Maybe something like this? arstechnica.com/gadgets/2024...
comment in response to
post
Thanks for the writeup! Is there more you can say about the "tuning" of o3? Was it specifically finetuned on the public dataset or was the public dataset just part of the training corpus? I guess the line here is a bit blurry.
comment in response to
post
Ah, interesting. If you implemented this with a "question asking LLM" and a separate "question answering LLM", would that make a difference? Maybe the "agentic" descriptor best applies on the on the system level.
comment in response to
post
It's true that the model can use that extra context to make the right decision, but I have an (unfounded!) intuition that it won't make a big difference. I think I have a mental model of LLMs very strictly following the prompt.
comment in response to
post
I don't think that changes much for me. I think a lot of my intuitions come from anthropomorphising the LLM. If you don't trust an employee's abilities you might give them specific instructions with exact criteria. A trusted employee is told "use your judgement to pick the best tool".
comment in response to
post
In code1 the LLM just has to make a classification, while in code3 the LLM has to "figure out" the criteria for choosing one option vs the other and then make a classification/decision based on that criteria, so we're trusting it do do more.
comment in response to
post
Yeah, I think the level of trust you're putting in the model is increasing across the code examples, which I think is an important part of the "agents" promise.
comment in response to
post
I don't have a great definition for "LLM-agentic", but to me it has to do with multi-turn interactions with external interfaces (especially interfaces with real-world consequences). None of these seem to point very far in that direction.
comment in response to
post
These articles sound like the settlement would be a one time payment of 100M. It should be the amount of revenue being collected from NJ drivers, minus the additional cost to the MTA of increased ridership. Probably around 200M per year.
comment in response to
post
This slide implies that the performance of ANNs can be attributed to their being brain-like, but maybe we could have gotten here by starting with completely different differentiable functions. But it seems like many neuroscientists here think they aren't sufficiently brain-like to be intelligent.
comment in response to
post
I found it notable and troubling when a bunch of right-wing people celebrated murder and I find it notable and troubling when left-wing people do it. The murders themselves are not that notable.
comment in response to
post
Whereas Matt believes in the "centrist" framework where fighting with people on the further end of your party makes you more favored among the general public
comment in response to
post
Here's my understanding of your key disagreements:
1) You believe that some of the policies that Matt sees as politically toxic could be sold better to the public
2) You believe that arguing with people "to your left" is politically costly and therefore Matt is doing the movement a disservice
comment in response to
post
So under this framework you find the policies that have positive cost/benefit in both politics and climate impact and pull those as hard as possible.
comment in response to
post
My understanding is yes, with the caveat that there are also political costs for pulling certain policy levers vs others. In his framework if you pull too many of those levers you don't get to do anything because the other guys are in charge. I think you disagree about these political costs.
comment in response to
post
I think Matt's style is really annoying to a lot of people (and I totally get it!). For what it's worth, Matt lays out the how he would personally evaluate the tradeoffs in the article: he would trust a scientific and economic consensus analysis for the global social cost of carbon.
comment in response to
post
I follow both of your work. It seems to me that you have very few substantive disagreements about the facts or your preferred policy (except that Matt supports investing in nuclear power). I think he would say that invoking existential risk is itself a way to avoid talking about tradeoffs.
comment in response to
post
The lightning connector was only USB 2 (60 MB/s) so this must only apply to devices with USB C
comment in response to
post
That's true. They really should have enabled watermarking. Unfortunately now that Llama 3 is out the cat is kind of out of the bag for homework cheating.
comment in response to
post
The potential upside of AI tutors really is huge though (see Bloom's 2 sigma problem). It seems like we'll be able to made progress on this in the next 10 years, but I haven't seen much in the way of actual useful implementations yet.