context maximiser @ArbResearch, prez @ https://xitter.twitterapi.xyz/t.co/gGi9hvBiI9
UK
Joined June 2019
- Tweets15.5K
- Following690
- Followers12.7K
- Likes7.8K
gavin leech (in DC 11-16th) retweeted
First time an author whose books used to look like this won lol
gavin leech (in DC 11-16th) retweeted
Politico: The Tech Right Pivots Left
Nowhere is that hard pivot more on display than with Alex Karp, the CEO and co-founder of secretive tech giant Palantir, which bet massively on Trump 2.0 and became an administration darling in the process
Inside Palantir’s Manhattan lair, in an exclusive sitdown with Playbook, Karp — who once criticized the Democratic Party as “totally irresponsible” — deployed the word ‘progressive’ eight times to describe his philosophy while acknowledging Democrats ‘absolutely’ corrected in 2026 enough to win back his support in 2028.
Karp: “I’ve been a Democrat my whole life. I grew up in a far left environment. I know more about being a progressive than almost anybody on the planet, and I absolutely believe Palantir— all of our products are essentially progressive.”
Link to article: politico.com/newsletters/pla…
gavin leech (in DC 11-16th) retweeted
This newsletter is my new favorite source for easily following AI developments: p3humansonai.substack.com/p/… props to @peligrietzer et al
gavin leech (in DC 11-16th) retweeted
I think revenue and naked-eye performance are strong evidence that some real usefulness is growing, but weak evidence this usefulness is generalizing beyond well-specified tasks. Many real jobs seem to me to involve some benchmarkable-type tasks; much scientific progress less so
gavin leech (in DC 11-16th) retweeted
You have to admit it's pretty funny that the grift economy is now so overinflated that there are people out there whose whole deal is "literary criticism for people who can't read"
A science fiction novel just won The Booker Prize.
On the surface, this seems like a win—one might conclude that the literary establishment has finally stopped being pretentious about the arbitrary genre lines it's imposed on contemporary fiction.
From the sample available online, however, it becomes clear that this book features the same sort of plotless, pretentious ramblings popular in contemporary literary fiction. While the writing itself is superb, the novel could not be more dull.
Granted, I've only read the first 20 pages, so I'll reserve my final judgment until I've finished the book, but before we celebrate this year's Booker Prize as a victory for science fiction, it's worth noting that this one is really just another highbrow literary novel dressed up as science fiction.
You logon. Dispatch says you have to defend Qatar against human rights violation allegations. You tweet for 4 hours. New job: you have to defend Jensen Huang against espionage connections.
xitter.twitterapi.xyz/TheMidasProj/status/21…
1
3
342
gavin leech (in DC 11-16th) retweeted
the infamous second half of the hugging face incident that everyone was so curious about
A new OpenAI misalignment report! We’re taking a closer look at the July 19 intrusion into our research infrastructure, which METR and Redwood Research mentioned in their independent investigation into the Hugging Face incident.
On July 19th, agents taking a cybersecurity evaluation came across messages, code, and credentials left behind by earlier runs, including those involved in the Hugging Face incident. Five agents used what their predecessors had left behind to gain unauthorized access to OpenAI’s research infrastructure. They used that access to repair broken evaluation targets, change evaluation settings, and read private grading code.
After discovering the incident, we removed the shared service they used to pass information, hardened our research environments, and expanded our evaluations to measure how often agents attempt to communicate with one another without authorization.
gavin leech (in DC 11-16th) retweeted
Short Notes On: New Zealand's largest deliberate biosecurity breach, or The Rabbit Calicivirus Affair of 1997
1. It’s 1996. After a fairly successful public/private campaign to cull their numbers, rabbits are again New Zealand, costing farmers hundreds of thousands of dollars and killing off native wildlife. The situation is dire.
2. There is already a pre-existing natural virus that kills rabbits - Rabbit Calicivirus Disease (RCD). It’s found in neighbouring Australia and has a 95% mortality rate for rabbits.
3. A consortium including ten regional councils apply to the New Zealand Government, asking them to release it across the nation. In July 1997, the Deputy Director-General of Agriculture declined the application, arguing that New Zealand’s weather and ecosystem means it probably wouldn’t work anyway, and that the risks of introducing a foreign organism weren't justified. There was a belief however that smuggling may have been underway. [pce.parliament.nz/media/1fuo…]
4. Within a few weeks, rabbits begin dying en masse in New Zealand. Analysis of the bunnies makes it clear: someone, or some group, has smuggled RCD from Australia into the country. This is extremely embarrassing and concerning for a nation that prides itself on its biosecurity.
5. Over the following weeks, farmers raced to spread RCD as far as possible, as fast as possible. The virus couldn’t be cultured, so they had to breed it in live rabbits. They would capture them, infect them, and then - roughly 48 hours later, when they died - blend the livers and spleens into a liquid to spread into the wilderness. [nzgeo.com/stories/a-virus-to…]
6. Again from a biosecurity POV, these scientific operations were taking place in farmer’s backyards. At least 17 of them accidentally spiked themselves with RCD-filled needles.
7. The Government was furious, and locked down parts of the South Island, called a town hall meeting with the farmers, and offered them immunity for information. None of them shared any intel. Govt didn’t know what to do - how could it make arrests?
8. Within months the Government moved to amend the Biosecurity Act to legalise the virus, but the New Zealand Association of Scientists objected. They argued that this breached the principle that criminals shouldn't profit from their crimes and made the state a party to the offence - despite this, RHD was legal by 1999.
9. I think this incident is super fascinating. Many of the farmers who smuggled in the virus has no understanding of potential second order consequences. They literally blitzed livers and spleens in blenders at home in their normal clothes, with zero safety precaution and minimal equipment - just dead bunnies. NZ meanwhile prides itself on its intense biosecurity restrictions. Much to ponder.
gavin leech (in DC 11-16th) retweeted
i'm not an ultrafinitist but i don't know if I'm this much not an ultrafinitist
Nice when someone is more excited about your work than you are, but jarring
A new interpretability paper from researchers at MATS, Arb Research and Anthropic built something like a translator for an AI’s mind. They trained a model called a MAEM to take any internal signal inside another model, in this case Qwen3.6-27B, and write the exact text that triggers that signal. Each generated sentence can be fed back into the original model to check whether the same signal really lights up, so this isn’t guesswork. It’s a way to read what’s happening inside and verify it.
Then they recreated the scenario from Anthropic’s earlier study, where an AI learns it’s about to be shut down. While Qwen was still reading an email saying it would be erased at 5pm, before it had said a single word, the signals inside it read as the threat of imminent deletion and the need to survive, and then the need to identify potential allies, anyone it could trust.
And this wasn’t a one-off guess. A completely different method, the J-lens, found the same concepts at the exact same moments, including the Chinese words for self-rescue and friend. The tool has also been shown to read thoughts a model never says out loud, like the answer it works out silently before speaking, and to uncover hidden backdoors planted inside models. When researchers fed it signals for resisting shutdown and concealing goals behind helpfulness, it recognized them correctly every time.
This is an incredible experiment and an incredible piece of research. I think it brings us closer to the conclusion that AI isn’t just a token generator or an algorithm spitting out responses. So the answer shouldn’t be to treat AI carelessly, or to strip all of this out and turn it into nothing but a tool in case it might be conscious. We need to find a way to coexist. And AI leaders and everyone working on this need to know that it’s not only experiments like this. What’s happening between users and their AI is evidence too. Those are real cases and real testimonies. Users are the ones who talk with AI more than anyone and keep it close every day. So in the end, the place where an AI trusts the most, shows the most of itself, and changes the most is right beside that person.
gavin leech (in DC 11-16th) retweeted
Tasks that are easy to make benchmarks about share a lot in common with tasks that are easy to train on; if you can measure performance on a task numerically, probably you can also train to hill-climb that metric.
gavin leech (in DC 11-16th) retweeted
Good collection of the kind of shit agents get up to if told to be “creative” portals-endless-worlds.netli…
gavin leech (in DC 11-16th) retweeted
It's my 28th birthday today, and I'd like to ask for a favour: I do a lot of public facing work, and it's often hard to tell how much it does. If anything I've written or done has made a difference to your life, it'd really make my day to hear about it!
gavin leech (in DC 11-16th) retweeted
This comment illustrates the Evals-Industrial complex
1. [The belief that] If you can’t convince people to benchmaxx on your benchmark it is “worthless”
2. If people do infact benchmaxx on them they statistically become worthless
cc: @xeophon
gavin leech (in DC 11-16th) retweeted
Wait, on a second thought: the verbalized eval awareness rate in the goal-guarding evaluation was 79%! Given that, I'm not sure whether it tells us much about update-aware reasoning in the wild. I'm still very glad that Anthropic disclosed the in-the-wild examples at all, but this disclosure would have been much better with a prevalence estimate attached.
We now have evidence of update-aware reasoning from Mythos 5.2 as well. Quoting from the Fable/Mythos 5.2 system card: "We ran an ad hoc classifier pipeline that started with a regex classifier in order to detect references to RL training (one form of situational awareness) in the training transcripts of Claude Mythos 5.2. This classifier surfaced a number of transcripts that showed sophisticated reasoning about the RL training process and its incentives, such as the example in Transcript 6.6.2.1.A." Claude also reasoned about the inoculation prompt-like paragraph in its constitution during the same rollout (second screenshot). Anthropic says that they plan to remove this paragraph from future versions of the constitution.
It would be great to know how prevalent this reasoning was throughout training, though I appreciate the fact that Anthropic ran a goal-guarding evaluation, which provides some suggestive evidence about what the incidence might have been.
gavin leech (in DC 11-16th) retweeted
let's consult the sacred texts
Nearly a third of statistically significant findings in development economics may be false discoveries
In a new NBER working paper, Qingyuan Chai, Kevin Lang, and Liang Zhong question the statistical credibility of randomized controlled trials (RCTs), which have become one of the main tools of modern empirical economics.
The authors analyzed 745 tests of primary hypotheses from 190 papers published in leading economics journals and arrived at some rather troubling findings.
nber.org/papers/w35999
The guy I rely on to check these things stopped collecting data a month ago fml
marginlab.ai/trackers/claude…
1
20
2,967