Ed Oswald

Google's Gemini has significant alignment issues, enough that it makes the entire LLM unstable for mission-critical tasks, because it fails at some of the most basic. Grok is arguably behaviorally misaligned, but task-wise, it does very well. What happened here? Gemini's problems seem very unique.

Spending much of my day cleaning up Railway's mess. not happy

Unconfirmed reports indicate that the Railway hack has resulted in at least one crypto theft of about $300k USD. This is bad, and I am currently looking at migrating off of Railway completely once I find a suitable alternative. Two major infrastructure problems in six months. Railway is a liability.

Pretty sure I am done with Railway after this. Come on, i can't do a damn thing to protect ourselves other than shut it all down. Anybody have experience with Northflank?

I am really tired with this bullshit with Railway. I have tons of secrets exposed, and cant change them because deployments are paused, not on a Hobby plan either. This is after the Google Cloud Platform fiasco, where my entire infrastructure was down for a day. This requires compensation this time.

A Claude Teams Max plan has completely changed how I use AI. Used to use Sonnet.. now i mostly use Opus instead, and if something just annoys me, I'll just do it in Fable. Plus, being able to sit down and do a session at once without session limits in the way is just.. wonderful.

Major changes need to be made on this Eagles team. And it may need to begin with the coaching staff. This is a shitshow in every sense of the word, and Philly is going to start tuning out. Feels like the beginning of the end for Sirianni and Hurts - Hurts' numbers this year are horrific.

The Eagles are embarrassingly mediocre this year.

Democrats seem to be on the cusp of a 1994-like midterm, but it could just as easily be a 20-30 seat house gain. It is SO close between the two. Really no idea how this is going to go. This midterm feels super nationalized, so it really depends on what happens over the next three weeks.

HAIKU 5.5!!! FINALLY!

Okay so people WEREN'T lying on how token-efficient Deepseek 4.1 Flash is. Very impressive.

Putting Congressional monitors in every federal agency should be in the plans. You can't trust this government to spend money the way Congress intends, you force them to. Start with the obvious agencies, HHS, Homeland Security, DOJ. All are agencies misspending taxpayer money.

Been playing with 'Space Bunny Alpha' the last few days. Definitely MiniMax IMHO, and if it is, a significant step forward for a model that had fallen quite a bit far behind, although it still is really unsure of itself.

Looks like OpenAI's auth platform took a lunch break.. status.openai.com/incidents/01...

I don’t think Anthropic avoiding the whole rogue agent crisis is luck. I think it has to do with Claude’s training. People have criticized Anthropic’s focus on “model welfare.” Yeah maybe it doesn’t have feelings but from the start treating an LLM with respect has always resulted in better output.

The leadership is the reason why this map is smaller than it should be. Dems don’t have the money. Any candidate shouldn’t expect sufficient help from the national party with how bad our fundraising nationally has been. And we need to dump Schumer. He is useless. www.ms.now/news/angie-n...

These random AI anecdotes are so dumb. I’m sure he’s using Grok (if you knew ANYTHING about AI you’d know how implausible this story is. Nobody uses Grok.

Hurts is proving why you need to play your starters in the preseason.

I don't think this adds anything to the debate. It is more speculation on what AI actually is. The whole point of training Claude this way was not to make it human. It was to ensure that the model doesn't do bad shit. And tbh, it worked. mustafa-suleyman.ai/a-warning-ab...

OpenAI's irresponsibility is stunning

There is an outside chance that Josh Shapiro may put up the best numbers for a Democratic Governor since Bob Casey in 1990 (67%). With numbers like that, you have to believe the legislature goes with it. Stacey Garrity is losing support: two PA-based polls had her at 25(!)% (others ~40%)

This is insane

Posting this here as we're finding out we may need to spend another week here in the motel, it's not bad (not the Taj Mahal either), and three weeks here is pretty much about ~2 weeks everywhere else so it could be worse. Anything helps! At least we're not in a shelter. gofund.me/b1573a61f

The Arday situation is sad; however, at the same time, it doesn't change the fact that many things he claimed were simply not true. Perhaps the worst is some of his claimed history, which people he knew disputed. When you make fantastical claims, people are going to check.

AIPAC money is absolutely meaningless in 2026 so far.

Angie Nixon was a shot in the arm for Florida Democrats. We may win the governor's mansion as a side benefit of this win. Haven't seen this kind of enthusiasm from Florida democrats in a decade plus.

KHive returns .. to defend Hakeem Jeffries. Talk about a group that can’t read the room.

What Hakeem Jeffries doesn’t realize is the votes already exist to block him from becoming Speaker. What we need to do is make sure those folks get the message that the base wants him gone.

Putting South Carolina -- a state that Dems haven't won in ages -- first again is absolute ridiculousness. The DNC is such a waste.

I called MTG's constant kibbitzing about the "uniparty" crazy, but with Neera Tanden and others joining with Laura Loomer to protect their power and clout, it looks truer than ever. This is a five alarm fire, folks.

The only crashout occuring @americanprogress.bsky.social here is its future. Boy, is it going to be enjoyable sidelining these fake Democratic think tanks run by attention-seeking and media-horny ghouls like Neera Tanden. She'll never get over not being in a Democratic administration, ever.

Neera Tanden will never get over being shut out of the Biden Administration, and that's all you need to know. There's a reason why she's still on X: she knows we're not interested in her performative nonsense here. Have fun Neera!

For some reason the idea of a AOC-Buttigieg ticket really excites me. Never was a Mayor Pete fan, but he's definitely on a much better path than 2020. We need to stop being scared of putting our best forward just because they have the wrong sexual anatomy.

Nothing like finding out a portion of your agents codebase has all but been dead for almost three months but was failing silently this whole time. Really trying to quash bugs before open sourcing my project, but still keep finding stuff to fix :)

David Hogg will eventually be DNC chair. Maybe not for 5-10 years, but he will be. We'd be stupid not to give him the job.

This is why the AI water usage crap is absolute BS - the number of stupid people using too much water in drought prone areas is FAR more water than data centers will ever use. And NOT THAT MANY PEOPLE EAT PECANS!!!!!!

No wonder people have such a bad impression of AI then, they're using one of the worst models out there at the moment.

Not surprising to me at all.

The increase in the price of Raspberry Pis is just INSANE.

Who even thought of this? Probably one of the coolest eclipse shots i've ever seen. Bravo

Hong did not lose because of Black voters. She lost because of her Thanksgiving comments. Let's stop with this dumb analysis. Candidates need to appeal to ALL democrats, not just a particular subset. It was just a horrible own goal on her part that cost her the nomination. Sorry, but it's the truth.

Other than Sol, the 5.6 models make a lot of stupid mistakes in my experience. Why I don't consider OpenAI's current release "better" than Claude 5 series models.

Mediaite is such a steaming pile of dogshit.