EB66

> I think most of it is bot driven spam by the various labs. It's hard not to notice 3 month old accounts with very strong opinions about various frontier labs and little else.

I've assumed the same as well.

I also assume that many of the companies developing these models engage in benchmaxxing.

At my company we've developed our own internal benchmarks for evaluating LLM models as they become available. The benchmarks are tailored to our particular use cases but the utility and knowledge our benchmarks assess is still fairly universally applicable. I see wide differences between what our internal benchmarks report and what the major benchmarks do.

There was a whole lot of fanfare about how amazing GLM 5.2 was when it was released, but it was pure rubbish on our internal benchmark -- far behind OpenAI, Anthropic, Gemini, DeepSeek, etc. I don't know how to reconcile the fact that GLM 5.2 performed very well on some of the major public benchmarks, but consistently performs so poorly on ours. OpenAI models tend to dominate our internal benchmarks.

That's interesting, we've had the same results as you at my company. gpt-5-mini was the clear pareto frontier leader for our in-house LLM benchmarks -- benchmarks that we built and tailored to our specific use cases. Then gpt-5.6-luna came along with the price cuts and immediately supplanted gpt-5-mini.

I always thought it was a little odd that gpt-5-mini was the leader for so long when more popular benchmarks placed gpt-5-mini further down the roster, but it seems you had the same result too.

> I would say GT is hardest to get in, but most inexpensive. While the later two are much easier.

Regarding the acceptance rates, it's actually the opposite. GT OMSCS has a very high acceptance rate (~80%) and UIUC and UT Austin online masters are much more selective (~30%).

It's great to see some hard numbers and analysis on a phenomenon I've experienced for quite some time. I use ChatGPT and Gemini as my two daily drivers. While I prefer the analysis capabilities of ChatGPT, I've been favoring Gemini in recent months because it tends to provide simpler, less verbose responses (at least with code) and over-edits to a much lesser degree.

If I need deeper analysis or a more critical second opinion, I'll go with ChatGPT. But for generating a quick simple changeset that I can then mold into what I need, Gemini is superior, and the data in this article reflects that.

Regardless of the model, my default prompts always include instructions to avoid unrelated refactors, favor the minimal change set that resolves a bug, etc.

The fact that Gemini returns the highest rate of fake positives aligns with my experience using the Gemini models. I use ChatGPT, Claude and Gemini regularly and Gemini is clearly the most sycophantic of the three. If I ask those three models to evaluate something or estimate odds of success, Gemini always comes back with the rosiest outlook.

I had been searching for a good benchmark that provided some empirical evidence of this sycophancy, but I hadn't found much. Measuring false positives when you ask the model to complete a detection related task may be a good way of doing that.

Sharding can be made mostly transparent, but it's not purely a DB-level concern in practice. Once data is split across nodes, join patterns, cross-shard transactions, global uniqueness, certain keys hit with a lot of traffic, etc matter a lot. Even if partitioning handles routing, the application's query patterns and its consistency/latency requirements can still force application-level changes.

I also had a close family member who committed suicide shortly after going on Prozac -- this also happened nearly 30 years ago. His young son later went on Prozac himself (several months after his fathers suicide) and immediately started demonstrating bizarre disinhibited anti-social behavior (e.g., damaging property, stealing from friends, etc). He was immediately yanked off Prozac when he started articulating his own thoughts of suicide. The bizarre anti-social behavior improved after discontinuing Prozac.

For some people, Prozac is a very dangerous drug. It is fully deserving of its FDA black label warning (which it didn't have 30 years ago).

You don't even need to film it, it's on Google Maps Street View: https://maps.app.goo.gl/cdyFttFsQPhpBHR48

That street view was filmed taken two months ago, when it was still warm and nice out, so tent activity would've been at its peak.

I wish I had seen your post earlier because I literally walked that stretch of road earlier this evening -- a couple friends of mine from out of town are going to the Patti Smith show at the Paramount tonight and we had drinks nearby. No tents in sight though we did encounter someone walking around with a blanket wrapped around their head. But still, one probable homeless drug addict is hardly "overrun".

It's really not as you describe... I agree things were getting worse in ~2019 and then became way worse during the pandemic, but it's much different now.

> I lived there for a very long time

It sounds like your last visit was during the COVID-19 pandemic. Homeless encampment conditions in downtown Seattle and throughout the city have much improved since then. Today, visible homelessness is effectively the same as it was back in ~2005.

I don't know why you're being downvoted. I've lived in the area for nearly 20 years and I agree that his description is far exaggerated. It was true in the depths of the COVID-19 pandemic (when any eviction or forcible displacement of people was prohibited), but not since then. Today, the visibility of homeless encampments in Seattle is roughly the same as it was 20 years ago.

I'm not saying that 37 out of 50 is good. I'm saying that 47 out of 50 is bad. Your data source doesn't refute the OPs argument that Arkansas is not a great place to live -- it actually supports the OP.

You are citing the US News "best states" ranking. In that ranking, California is ranked #37 overall and Arkansas is ranked #47 overall. Even your own hand picked data source supports the OP...

Yeah, very true, when we're hiring we'll sometimes skip over the most talented candidate and pick a candidate who we feel would be the better communicator, easier to work with and fit in with the team. Easier said than done, but we at least try to hire for that.

This is a really interesting phenomenon that I've experienced before myself but I hadn't fully understood or appreciated as clearly as the author has.

At my company we put a big emphasis on code reviews. We encourage devs to pull request code fairly regularly to keep PRs relatively small (when possible) -- before so much code has been written that it's not really possible to change course without blowing up deadlines. We encourage our junior devs (who might not be capable of identifying bugs or proposing fixes on code written by a senior) to ask questions in their code reviews -- to verify assumptions, to request an explanation of how something works, confirm that a particular edge case has already been considered, etc. It can be hard to get a junior dev comfortable with doing this (questioning a senior dev), but even if the junior isn't identifying bugs it will often lead the senior to better understand their own code and the architectural concepts that underpin their own coding decisions. Like the author points out, this only happens because the senior dev endeavors to explain their work to the junior dev (Protege effect). Also, a good many times it leads the senior dev to re-consider how they wrote something and they might add a revision to address a possible edge case not previously considered. I hadn't thought of it this way before, but this is the Socratic method that the author talks about.

We also put a big emphasis on in-code comment writing -- largely following the commenting principles laid out by John Ousterhout's "A Philosophy of Software Design". These comments are of course for long-term maintenance purposes, but they also benefit team learning. Class, method and variable naming are obviously important too. Our internal code reviewing mantra is that 'I want to be able to read your code like a story book -- when I get to the end, I at least want to be able to understand what happened'. Not always possible, but a good goal. Writing comments and choosing class/method/variable names in pursuit of that goal massively contributes to the learning of the team. During our code reviews, one of the most common requests by reviewers is for the author to add a comment explaining something that was very difficult (or impossible) for the reviewer to grasp on a first read.

This approach has worked very well for us. Everyone learns and our product quality improves.

> because beyond simple single-purpose applications with flexible uptime requirements, that’s not a realistic option.

I frequently hear this point expressed in cloud vs colo debates. The notion that you can't achieve high availability with simple colo deploys is just nonsense.

Two colo deploys in two geographically distinct datacenters, two active physical servers with identical builds (RAIDed drives, dual NICs, A+B power) in both datacenters, a third server racked up just sitting as a cold spare, pick your favorite container orchestration scheme, rig up your database replication, script the database failover activation process, add HAProxy (or use whatever built-in scheme your orchestration system offers), sprinkle in a cloud service for DNS load balancing/failover (Cloudflare or AWS Route 53), automate and store backups off-site and you're done.

Yes it's a lot of work, but so is configuring a similar level of redundancy and high availability in AWS. I've done it both ways and I prefer the bare metal colo approach. With colo you get vastly more bang for your buck and when things go wrong, you have a greater ability to get hands on, understand exactly what's going on and fix it immediately.

> Is flying in a thin atmosphere harder?

Yep. Setting maths and motors aside... it's why helicopters have service ceilings and why no one has built a helicopter that can fly away into the stratosphere.

I'm not interested in the competition rules and criteria. I'm just curious why the OP views Impetus as "better".

Impetus is certainly more artistic and thought-provoking, but is there anything that makes Impetus a greater technical achievement than Street Food? From my (very) non-expert perspective, Street Food seems much more technically impressive.

> Where people come up with that kind of cash is beyond me, but they're coming up with it one way or another.

A "cash buyer" just means that the buyer didn't open a new mortgage to close the sale. Cash buyers often bring in cash from other interest-accruing loans -- such as HELOCs or portfolio loans. Foreigners or irregular income earners are a couple examples of people who might not qualify for a conventional mortgage and would need to tap into alternative loans to buy a house and thereby become a "cash buyer". Cash offers are also considered more competitive (they close faster), so someone might make a cash offer (via other loans) just to make their offer have a higher chance of succeeding.

> Sorry, just because you found this trivially fixable array mismatch doesn't mean your buggy patch has to be accepted the way it is.

That's not really what anyone is suggesting should have happened. The maintainer should've just taken the two minutes of extra time required to checkout the OP's code, make whatever changes the maintainer wanted and commit it with co-authorship.

That clearly would've been the best course of action for the maintainer to follow (especially given that the OP expressed his desire to the recognized for his contribution) -- the best fix gets merged in and the work of others is acknowledged/respected. That said, I'm not surprised it didn't happen -- social niceties often seem to escape the maintainers of these sorts of repos...

For anyone contending with the menace of fruit flies or gnats (and especially if you own a fair number of indoor plants), I could not recommend these more wholeheartedly:

https://www.amazon.com/Stingmon-Pack-Sticky-Fruit-Fungus/dp/...

It's essentially a better looking version of fly paper. They're easy to place alongside indoor plants. Fruit flies and gnats are attracted to decaying organic matter and the moist soil of indoor plants is an often overlooked oasis for them. Gnats are more likely to hang out in indoor potted soils (especially if overwatered) because they feed on fungus, but fruit flies can take refuge there too.

A couple years back I had a months long battle with fruit flies. Kitchen was always clean, garage sealed up tight, no indoor compost, I used apple vinegar traps of all styles but nothing solved the problem until I placed those sticky paper inserts at the base of all my indoor plants. Within a week the fruit flies were gone.

> Comments like this are really bizarre. It's like asking for "data" that says looking both ways before crossing the street is a good idea and reduces the frequency of pedestrian fatalities.

Since when is it "bizarre" to ask someone if they've based their assertion off of hard data? I replied to at-w's fairly remarkable claim that white collar criminals are less likely to harm people while out on bail than non-white collar criminals. That assertion is not a forgone conclusion like the wisdom of "look both ways before you cross the street". Clearly it's reasonable to ask the poster if they're basing their assertion off of some hard data.

> Those two people are asshole fraudsters, but they don't present a physical danger to people around them.

We're talking about the propensity to harm others while out on bail. Harm is not limited to physical harm. The two examples I gave were of white collar criminals who continued to defraud/extort others while they were out on bail.

> are objectively less likely to reoffend while on bail in a way that would harm regular citizens

I'm not so sure that's true... do you have data to support that? I can think of some counter examples pretty easily off the top of my head -- Billy McFarland, Michael Avenatti, etc.

> I'm dumbfounded that anyone knowingly moving there would be dumbfounded that things move at a much slower pace.

That's a pretty rude response and also invalid. He's not dumbfounded at the slower pace of life. He's dumbfounded that a nation of 300k could have a week long electricity outage and not have a single resident capable of fixing the issue or a backup plan ready to activate. It suggests exceptionally poor planning -- even for an island nation in the Pacific.

> These fears ultimately proved unfounded, since corium began dripping harmlessly into the flooded bubbler pools before the water could be removed.

That line in the Wikipedia article seems to conflict with the book I read by Andrew Leatherbarrow "Chernobyl 01:23:40". In the book he states that if the water hadn't been drained and if the molten core had reached the water it "would have done unimaginable damage and destroyed the entire power station, including the three other reactors."

It's a shame that the Wikipedia article doesn't cite it's source on that claim. It would be interesting to reconcile the seemingly conflicting information.

> Furthermore, the paper you linked was talking about wind directions not about whether the molten rods reached the cooling water underneath the reactor room floor.

I did not say that the IAEA paper was about molten rods hitting the coolant water. There is more than one way that Chernobyl could have been much much worse. According to the paper, different wind and rainfall could have made the disaster 200-400x worse in terms of radiation consequences to humans. To say that is significantly worse would be an understatement.

> Popular books have the incentive to garner sales - often through dramatization - not report accurate findings.

It seems like you've made up your mind and no amount of evidence to the contrary will change it. I don't think I'll bother to continue participating in this line of discussion.

You are certainly entitled to your own opinion, but others who have researched Chernobyl far more extensively than you or I have concluded differently. I am primarily getting my information from Andrew Leatherbarrow's book "Chernobyl 01:23:40". In that book, he states essentially the same as I have above.

However, it's not just Andrew Leatherbarrow who thinks so. All it takes is a simple Google search and you'll find reputable research papers hosted by the IAEA whose primary conclusion is that the accident could have been much worse. Here's one example: https://inis.iaea.org/search/search.aspx?orig_q=RN:18009127

I'm getting my information from the book written by Andrew Leatherbarrow "Chernobyl 01:23:40" -- not a TV series. Yes, according to him they are still alive. I said it was a probable suicide mission because all of them went in there knowing full well they might not make it out.

Yes, it did come close.

If it wasn't close, as you suggest, then why did a crew of people wade into highly radioactive water -- a probable suicide mission -- to manually operate the valves to drain the water?