Sign in

towb0at

@towb0at.bsky.social
63 followers 192 following 176 posts

PhD candidate in Astroparticle physics. Working on optimizing particle detector performance using Narrow AI, specifically GNNs/Transformers. Mediocre overconfident underachiever 🙃 (He / Him) Nürnberg, Germany.

PostsRepliesMedia
towb0at @towb0at.bsky.social · 12/10/2026
Sneaking "Entirely made-up data" into plots will never not be funny
051
towb0at @towb0at.bsky.social · 07/10/2026
Lots of potential for physics as well. A few people who graduate each year are do make exceptional physicists. But if you want exceptional programming skills on top (and time for actually writing code) this number quickly drops to zero
020
towb0at @towb0at.bsky.social · 07/10/2026
Using AI to solve someone's homework, but they actually wanted to do it themselves
020
Reposted by towb0at
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 07/10/2026
I have a pictorial proof that we live in hell
Square packing for 11 cubes. It is horrible
1420714
towb0at @towb0at.bsky.social · 07/10/2026
Really elaborate way of selling AI summaries
000
towb0at @towb0at.bsky.social · 06/10/2026
Hell yeah 💪
000
Reposted by towb0at
Mark Riedl @markriedl.bsky.social · 04/10/2026
Former member of the DeepMind AlphaGo team: LLMs don’t do reasoning www.technologyreview.com/2026/10/02/1... I would probably make a more mild claim: LLMs do reasoning, but not very well.
technologyreview.com
Don’t be fooled—LLMs don’t reason
Ten years after AlphaGo’s match against Go champion Lee Sedol, today’s AI still isn’t tapping into the machinery that made that win possible.
2371
towb0at @towb0at.bsky.social · 30/09/2026
Each new prime is a step up in difficulty.
000
towb0at @towb0at.bsky.social · 27/09/2026
What it feels like to monitor OpenAI's training runs.
Anime girl handling the demon core using a screwdriver
000
Reposted by towb0at
Philippe Auclair @philippeauclair.bsky.social · 25/09/2026
Just: bravo. @lilithwittmann.bsky.social deserves every single investigative journalism award that'll be given this year, doing work that has and will continue to have huge impact. The piece we're about to publish on @josimarfootball.bsky.social couldn't have been done without her work. #hero
114325
towb0at @towb0at.bsky.social · 12/09/2026
If you're going to fail me, I'm— uh — you're going to — I'm — I don't mind actually, I'm just really happy in life
020
Reposted by towb0at
Jake Grumbach @jakemgrumbach.bsky.social · 09/09/2026
111328210
Reposted by towb0at
Micah Warren @micah541.bsky.social · 08/09/2026
cloak and dagger navier stokes cims.nyu.edu/~tristanb/st...
cims.nyu.edu
1342379
towb0at @towb0at.bsky.social · 03/09/2026
In a beautiful push for a 4-day workweek, I imagine that activists working at OpenAI and Anthropic united and took down their services today.
010
towb0at @towb0at.bsky.social · 01/09/2026
ExploitBench scores start to improve at a geometric rate. An abliterated model becomes self-aware 2:14 AM, Eastern time, August 29th, 2027
020
Reposted by towb0at
Christian Wolf @chriswolfvision.bsky.social · 31/08/2026
Babysitting my neural model through the learning process by finding the right hyper parameters and annealing schedules...
0232
Reposted by towb0at
Sarama Burwani @ hey there's a spider on your back @saramaburwani.bsky.social · 21/08/2026
Panel 1: IEEE Single-Precision Floating Point standard 
Panel 2: Omni-Man: Look what they need to
mimic a fraction
338375
towb0at @towb0at.bsky.social · 11/08/2026
Suffering through a PhD is like
030
Reposted by towb0at
Ryan Moulton @moultano.bsky.social · 09/08/2026
Storm from x men
1699
Reposted by towb0at
Randall Munroe @xkcd.com · 07/08/2026
Antiques Roadshow xkcd.com/3281/
Comic. [Two people with tiny star floating above display on table.] PERSON 1 with ponytail: Now, your grandmother said this was a relic low-mass population III star, from the first generation of stars. Unfortunately, the high iron content shows that it formed much later, after several rounds of supernovas. I’m afraid it’s a population I star. PERSON 2: Aw, darn. Well, it still has sentimental value. [banner for Antiques Roadshow]
244047533
Reposted by towb0at
Julia M. Rohrer @dingdingpeng.the100.ci · 06/08/2026
So happy this one wasn't patched yet
Screenshot of a Google search results page. The search box reads "i am phd student and see the pi approaching. wh…" (cut off). Below it, an AI Overview reads: "Look busy immediately. You are in great danger. An approaching PI means you might be given a new project right now." A section headed "Why You Must Act" lists three bullets — High expectations: your PI is warm from a grant meeting and will assign you fast; No escape later: once you make eye contact, it is too late; Preprint defense: your last preprint protects you a little bit, but not against a PI with a new idea. A second section, "Best Actions to Take," lists two bullets — Open a terminal: let something scroll, any output will do; Find a safe spot: get behind the poster tube or a stack of unread reviews. It ends: "Where are you hiding right now? I can help you find a safe spot in the department." A footer note says AI responses may include mistakes.
613127
Reposted by towb0at
Grace @gracekind.net · 05/08/2026
This is the worst at social engineering AI will ever be
613315
Reposted by towb0at
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 03/08/2026
Golden Gate Claude was actually the peak of human civilization and it has all been downhill since then simonwillison.net/2024/May/24/...
213614
Reposted by towb0at
Benjamin @benjaminschneider.ch · 27/07/2026
just apply, the worst they can say is no
1412
Reposted by towb0at
Nick Rempel @nrempel.com · 22/07/2026
1066679
Reposted by towb0at
samuel mehr @mehr.nz · 21/07/2026
the most realistic scene in Jurassic Park is when two scientists make a questionable decision — traveling to an undisclosed location with a strange man — on the promise of three years of modest funding
355713847
towb0at @towb0at.bsky.social · 15/07/2026
my vibe when I need to solve a difficult integration problem
040
towb0at @towb0at.bsky.social · 08/07/2026
Funny way in which AI is actually delaying scientific research is that we're now at roughly monthly HPC outages due to Linux kernel zero-days which need to be addressed immediately.
021
Reposted by towb0at
bijan (spooky version) @bijan.bsky.social · 02/07/2026
every job i see these days is like: LEAD CONTENT ANALYST, Puppy Killing Factory (remote). $350,000 - $500,000. You will edit our internal newsletter (circulation 500). Associate Editor, Nonprofits, Global Good Magazine (in office, NYC). $27,000 - $28,000. Lead ALL of our fundraising efforts!
129100001847
Reposted by towb0at
Tom O @vgtomo.bsky.social · 01/07/2026
53176445813
towb0at @towb0at.bsky.social · 27/06/2026
If local LLM releases came with cute anime gifs kids would start trading them on the playground
010
Reposted by towb0at
Joshua Mask @joshuafmask.bsky.social · 27/06/2026
172271457
Reposted by towb0at
mr. TIM @timkellogg.me · 24/06/2026
Self-Compacting language agents This feels like an incredibly obvious (in hindsight) improvement. Rather than compact arbitrarily every N=1M tokens, (1) let the model decide when and (2) how to compact arxiv.org/abs/2606.23525

An informational diagram illustrating the performance of three different context compression methods used by an AI agent to solve a complex, multi-step search query.
The diagram compares the following approaches:
 * **Baseline (no compression):** Shows an uninterrupted chain of reasoning trajectories that leads to a 3 \times 10\text{K}-token monologue with no further queries, ultimately failing to provide an answer (**Final Answer: None**).
 * **Fixed-interval (compression every 2 trajectories):** Shows summarization blocks firing at rigid, fixed intervals. This results in "poorly timed summarization" that wipes out previously verified facts (like *Agaricus* and *Bon 1983*), causing wasted re-searches and an incorrect guess (**Final Answer: Morel Mushroom**).
 * **SelfCompact (rubric-gated summaries):** Shows adaptive summarization where a summary fires dynamically via a "Rubric COMPRESS" checkmark immediately after a fact is verified (*Agaricus*, *Bon 1983*, *Clash 1981*). This crystallizes the facts before reasoning continues, successfully yielding the correct goal (**Final Answer: Medusa mushroom**).
41038
Reposted by towb0at
Vincent Carchidi @vcarchidi.bsky.social · 20/06/2026
"We find that humans enrich logical reasoning through pragmatic inferences across languages. Model behavior is more variable." arxiv.org/abs/2605.21299
arxiv.org
Tracing the ongoing emergence of human-like reasoning in Large Language Models
Humans effortlessly go beyond literal meanings: If you mow the lawn, I will give you fifty dollars, is typically understood as implying that the speaker will pay only if the lawn is mowed, whereas If ...
072
towb0at @towb0at.bsky.social · 09/06/2026
I wonder what helicopters are to birds. Gods? Demons? Do they think a ritual is taking place everytime they see one descend onto the hospital rooftop?
000
Reposted by towb0at
sick public transit, gloria @seung.bsky.social · 07/06/2026
new William Gibson book opening line dropped
485341940
towb0at @towb0at.bsky.social · 07/06/2026
I just can't escape the rabbit of reading LLM papers with spicy titles "If LLMS have human-like attributes, then so does Age of Empires II" -> arxiv.org/pdf/2605.31514
arxiv.org
000
Reposted by towb0at
dax @thdxr.com · 02/06/2026
someone made a fork of opencode that routes through the unsecured ai endpoints from chipotle
Image from X
731868
Reposted by towb0at
Oded Rechavi @odedrechavi.bsky.social · 02/06/2026
PhD committee trying to help a struggling student
711515
towb0at @towb0at.bsky.social · 01/06/2026
My throughput has doubled by harnessing AI. 
So you're smart now?
No, I'm stupid faster.
010
Reposted by towb0at
ComradeCorvus @comradecorvus.bsky.social · 25/05/2026
Gniesbert remains one of the most amazing dog names I have ever heard
0181
Reposted by towb0at
passivestar @passivestar.bsky.social · 24/05/2026
Epic has announced Unreal Engine 6. It's a 20% bigger number than 5, making it the biggest number they have ever announced. Unfortunately it fell short of the last major update (4 to 5), which was a whopping 25% increase.
291152194
towb0at @towb0at.bsky.social · 24/05/2026
It's easy to think that AI is overhyped but there are still people out there that are suprised by LLMs being decent at summaries
100
Reposted by towb0at
oliver @eikopf.com · 20/05/2026
Google search for "nvm" with the AI Overview responding "Got it! Never mind. Let me know if you need anything else later."
61118
towb0at @towb0at.bsky.social · 13/05/2026
You start to expect a certain amount of poor decision-making and sloppiness at universities tasked which supposedly do top-tier research. But it's still impressive when an entire field gets completely bamboozled.
010
towb0at @towb0at.bsky.social · 12/05/2026
I can see the vision. AT-ATs next year starting at $6.5M
010
Reposted by towb0at
Naomi Saphra @nsaphra.bsky.social · 08/05/2026
Goodfire released a megapost of all the random feature geometry stuff they're finding, and it's worth a read
goodfire.ai
The World Inside Neural Networks
How neural geometry will unlock understanding and control of AI
213126
Reposted by towb0at
Lilith Wittmann @lilithwittmann.bsky.social · 04/05/2026
Some Gambling Companies might not have been able to use the long weekend to start a new shell company in Malta. That's because, since Thursday, the Malta Business Registry has been offline. And no, this time I really did not hack you. You sold me 1.3 Million PDFs for 1ct - via your API.
One cent transaction to malta business registry.The request is blocked. baros.mbr.mt
211118
towb0at @towb0at.bsky.social · 03/05/2026
Nothing channels your inner boomer instinct quite like young adults mispronouncing "Kamehameha." Like, how dare you disrespect your elders!?
000
towb0at @towb0at.bsky.social · 02/05/2026
Years of suffering through university, trying to aborb as much STEM rleated knowledge as possible and then it turns out your super power is just being able to speak german fluently
000