matthew robertson

Doctors in China have been executing prisoners by extracting their hearts for transplantation, according to a new paper by myself and co-author Dr. Jacob Lavee in the American Journal of Transplantation. we do a little thread about it. onlinelibrary.wiley.com/doi/full/10.11…

@ryancbriggs counterpoint: this is, in fact, a right wing belief

One of my more left wing beliefs is that we should be teaching more actual Marxism to undergrads. The “Marxist”stuff people complain about isn’t Marxist at all. Marx was wrong in important ways, but his ideas remain interesting and fertile.

@cremieuxrecueil @alexolegimas Pangram actually works. try it with many paragraphs and experiment with LLM text & real text. I was genuinely surprised. Their methods papers are also very clever. I'm still actually very surprised it works so well.

This flagged as 100% AI-written: "In contemporary discussions of technological change, it is increasingly evident that societal adaptation depends not merely on the availability of innovative tools but on the collective willingness to reinterpret long-standing assumptions about efficiency, equity, and human agency. While some observers emphasize the disruptive potential of automation, others highlight its capacity to catalyze novel forms of collaboration that were previously inconceivable. Taken together, these dynamics illustrate a broader transition toward systems that are simultaneously more complex and more interconnected, requiring a level of coordination that challenges traditional institutional boundaries." It also flagged a paragraph from an article I wrote as human-written (correctly), and two paragraphs from a draft on a corporate investigation as AI-written (incorrectly) when they were transcribed from handwriting. Then I ran out of credits for today.

@DavidDecosimo @UNC Got a syllabus up?

I'm teaching "Political Theology" @UNC next spring. I don't think there's been a course like it here in the last 125 years or so, if ever. It should be good. If you want to think & talk about politics, God, nationalism, justice, common good, & the anti-Christ, you should take it.

@thsottiaux Scrollback still token broken in zellij, sadly, making it severely less usable :(

Codex CLI 0.48 is here. Packs many quality of life improvements, MCP features, stabler /compact and auto-compaction, better documentation, --add-dir for additional working directories. Lots of contributions from the community, thank you. Full notes: github.com/openai/codex/r…

@richkuo7 @ccccjjjjeeee @mehulsharmamat @mehulmpt atlassian already has huge user base and ways to get new product in front of them...? couldn't they just recreate dia or arc or anything even 10x better with far less than 600m?? The explanation still doesn't make sense, in other words

@ccccjjjjeeee @mehulsharmamat @mehulmpt I suppose it’s relative. The ai browser market is very small

I have never understood how these acquisitions make sense. A browser with almost no market share, no monetisation, highly competitive market, for $610M all cash? Please educate me why (genuinely)

@gbrl_dick These all your shots dude ?

also probably use a camera with auto exposure and a good matrix meter — i’ve shot lots of slide film in leicas, which have no meter or basic spot meters, and i always cook like 60% of the shots and feel like an idiot

ektachrome 100 was reintroduced pretty recently and has become a film stock for motion pictures, so will probably stay around, but i think fuji makes the chemicals for developing e6 it’s a bit more blue than provia, and less sharp, but has nice blacks and slightly more latitude

*sniff* *pulls shirt* You know, this is perfect - *gestures wildly* - this is the ultimate perversity of capitalism at its purest. Here we have Anthropic, this company claiming to build "AI for humanity," and what do they do? They create this digital cocaine, this Claude Code, and then - *sniff* - they act surprised when people become addicted to it! "Oh, some users are running it 24/7" - but of course! This is not a bug, this is a feature! They created the perfect productivity drug, the ultimate tool of self-exploitation under late capitalism, where the worker voluntarily chains himself to the machine, coding through the night, becoming one with the algorithm. *touches nose* And then comes the beautiful part - *laughs darkly* - "one user consumed tens of thousands in model usage on a $200 plan." My God! This is pure comedy! They are shocked - SHOCKED! - that someone would extract surplus value from their system. But isn't this exactly what capitalism teaches us? To maximize profit, to exploit every inefficiency? This user is the perfect neoliberal subject! He understood the game better than Anthropic themselves! *waves hands frantically* But no, no, no - now they must introduce "limits." You see the perversity? First, they create infinite desire - "Claude will code for you forever!" - and then they introduce scarcity. This is how capitalism functions: it promises unlimited jouissance, unlimited enjoyment, but always pulls it back at the last moment. "You can have it all... but not really." And the most obscene part - *sniff* - they frame this as caring for the community! "This impacts capacity for all Claude users." This is the same logic as ecological austerity - we must all suffer equally for the common good, while the real problem is the system itself! They created a monster and now blame the users for feeding it! *adjusts shirt* You know what this really is? It's the perfect metaphor for our predicament with AI. We create these systems that promise to liberate us from work, but instead, they become new masters. The user running Claude Code 24/7 - he is not free, he is the most enslaved! He has internalized the demand for infinite productivity so completely that he needs an AI to keep up with his own superego's demands! *sniff* And so on...

Italo Disco is the soundtrack of the summer

@lbeurerkellner The github mcp is indeed at fault if the default setting is full access to private repos, no?

Regarding mitigations, we don't see GitHub MCP at fault here. Rather, we advise for two key patterns: (1) restrict agent permissions as much as possible (2) continuously audit and monitor your agent and MCP connections More details in the blog: #mitigations" target="_blank" rel="nofollow noopener">invariantlabs.ai/blog/mcp-githu…

😈 BEWARE: Claude 4 + GitHub MCP will leak your private GitHub repositories, no questions asked. We discovered a new attack on agents using GitHub’s official MCP server, which can be exploited by attackers to access your private repositories. creds to @marco_milanta (1/n) 👇

if you’re wondering what life was like 20 years ago without any major technological advancements, just travel to germany

work like this (among a few other reasons, probably) is why we shouldn't defund academia

Soviet Children living in Siberia getting UV light exposure during the long dark winter months 1987

Tinkering with human embryos will certainly be worth the risk.

@zetalyrae lol just read on hn but didn’t realise it was you

I wrote about what I use Claude for, and how my usage has changed since the 3.6 release

@nikitabier @HeinrichKuttler @pull_gs So phone companies are just reading everyone’s SMSs like that? That’s what the STOP trick implies.

@HeinrichKuttler @pull_gs It is required by the provider if you want to be compliant.

Lots of wrong answers in the replies. As the former King of SMS, it’s simple: It’s to warm the number. You cannot send bulk texts containing a URL until your number has earned credibility with phone carriers. To earn credibility, it needs to have replies from other numbers. So the message they send is always something that tries to elicit a response from you. Once response rate is high and the phone number is warmed, they send tens of thousands of spam links to people. Best way to fight back? Replying STOP in all caps will immediately flag the number.

@gbrl_dick feeling some 'record rip, and at that moment i was radicalised' vibes upon watching this clip. Just yikes

incredible how bad these people are at their jobs easiest answer is just: “a new one bedroom house in sydney is $x00,000, that’s obviously not sustainable. our goal is to build a lot of houses, and our hope is that the entry level price for new buyers goes down”

@gbrl_dick v7labs.com/go looks pretty rad

(UI)

are there examples of products that use LLMs without using a chatbot or IDE UX?

@DefenderOfBasic this is tailscale funnels you're describing

is someone going to build this or do I have to do it myself? (almost guaranteed this will go viral & should be extremely easy. Literally just a DB that redirects to an IP address)

@BenjaminDEKR they cranked temp to 1 tho right? That could be a lot to do with it?

Spent all weekend coding with o1-preview / o1-mini. My honest evaluation: It's just bad. More specifically: It's bad for actually building things in the real world. It's frustrating, and hard to precisely explain why. I'll try: The o1 model is far too eager to "figure out the problem" behind the scenes and in the process, doesn't work with you to actually fix the problem in front of you. It doesn't want you, the human, in the loop. It does its chain of thought under the hood, and then info-dumps what it thinks is the solution -- with piles of output to the point that it's information overload. This is fine if the proposed solution is exactly right. But in the vast majority of cases, it isn't. And then you have to go back, not just to the last output but somewhere in the middle of a giant stack of output. This becomes tiresome fast, and isn't actually very effective at iteratively fixing or improving code. Instead of, A -> B -> C "wait, let's go back and adjust B please" it is A -> B (B1, B2, another idea bout B2, B3.1, B3.2, B3.3) -> C C C C -> D (D1, D2) "wait, let's go back and.... ahh.... jesus christ..... stop.... no....." To put this in simple terms, it's like trying to methodically work on a car with an over-eager, rambunctious assistant who has ADHD and had too much coffee. They're smart, and they're trying to help... ...but instead of bringing you one tool at a time, they keep running out and grabbing a dozen different tools and the kitchen sink, trying to anticipate everything and just bringing confusion. "Oh we'll need to do this! And then this! Here's five manuals about that part, OH here's this other thing, have you considered this?!" It's this strange outcome where the model is technically very capable and smart, but in practicality goes off the rails when used to actually build. Just my honest two cents. My opinion may change, but keeps coming back to this.

@mervenoyann but this won't work for creating & searching a database of millions of files right? One still needs to create an inverted index and all that stuff for bm25. Not clear how this solves those problems??

I was giving consultancy about multimodal RAG to some external people, wanted to post it here stop using OCR + LLMs. if you want to retrieve, use ColPali, if you want RAG from docs, use vision language models