Palisade interviewed 22 current and former employees from OpenAI, DeepMind, and Anthropic about their personal views and fears around AI development.
Today, we’re releasing the first batch of those interviews.
Please watch and share.
One perspective from inside Anthropic:
Q: You said something like people would be right to be angry at the people developing AI … but you're a guy developing AI! Should people be angry at you?
Tao Lin: Yeah.
Yeah, people should be angry at me and everyone else developing AI.
It's really not that bad to have people be angry at you.
You know, we live in a bubble. I won't meet those people.
Q: This is somewhat hard to understand. Like, do you think that you're doing something wrong?
Tao Lin: I don't currently think I'm doing something wrong.
Watch our new From Inside AI interview with current Anthropic employee Tao Lin. If you haven't talked to a lot of AI company employees, you're likely to find a lot that will surprise you here
Q: What's it like to talk with the other people who build AI about these kinds of things?
Tao Lin: I think I've become very desensitized to it all when I'm talking to my co-workers. I think I've been talking about risk from AI with my co-workers for the last three years or for the past six years.
Tao Lin: It's just every day for us.
Tao Lin: I think that there was a vibe shift or change in priorities in the past few months. I think that this was primarily caused by the OpenAI/Hugging Face incident and secondarily by Jacob Coxon departing Anthropic and we've invested more in security and more in safety since then.
Q: Are those changes in investments going to solve the problem?
Tao Lin: No
Palisade Research's @JeffLadish says AI progress is accelerating faster than even the AI 2027 scenario predicted, with Anthropic researchers expecting most AI development to be automated by mid-2027:
"It's really notable that this is a faster speed up than Daniel Kokotajlo's AI 2027 scenario predicts."
"A year ago, a lot of people were sort of down on Daniel Kokotajlo cause he updated a little bit towards longer timelines than as predicted in AI 2027. And now he’s saying, ‘No, I think it actually might be shorter timelines than we predicted.’"
"That's insanely fast progress for something that was already going really fast."
"I talked to a lot of people at Anthropic, and they say, 'No, I think in early to mid 2027, we'll probably be able to do this.' Most of the AI development will be happening autonomously by AI agents. That would transform the world so fast."
"It's just easy to be within a company and see a lot of progress and sort of drink your own Kool-Aid."
@PalisadeAI
METR says AI progress has accelerated 2x. But the strongest evidence is still redacted.
Palisade Research's @JeffLadish argues Anthropic should release the data so the world can assess the findings for itself:
"I just take METR's reports really seriously. I think they have a really good track record at being extremely rigorous."
"When they say, 'Here's the situation. It looks like a 2x speed up to us than previous trends based on AI uplift,’ I’m like, that’s pretty good evidence."
"METR said that they had to redact most of the best evidence for this in the report. They're basically saying, 'This is our assessment, but if you don't trust us, you shouldn't trust this assessment. Anthropic is not sharing the data with you.’ And I’m like, this seems like not good."
"I trust METR, but I don’t think that the world should have to trust METR on this one."
"Anthropic should release this data, un-redact this data so that the whole world can assess this for themselves. This has huge strategic implications for national security, for US-China stuff."
@PalisadeAI
Palisade Research's @JeffLadish says AI models are getting better at almost everything, but still struggle with humor, charisma, and persuasion:
"When I use Opus 5.5, I’m like, 'Wow, this model is better at everything,' or most everything. Anecdotally, I feel like I see some generalization with every model. But it sure is spiky."
"One of the things I'm surprised by is that the models are still not very funny or charismatic. I was expecting to see more social impacts from models just being better at modeling humans and manipulating humans for better or for worse. And that’s not what we see."
"If you look at where most of the humor and the persuasion is coming from, I still think that’s mostly coming from humans modeling humans and steering the models to do that."
"When AI companies really seem to push on a particular capability and build RL environments for it, I don’t know of any instances where they just totally fail. It seems like they mostly succeed."
"It just looks like the models wanna learn and that if you just put them in the right learning environment, they’ll learn the stuff."
@PalisadeAI
Forget the 40 AI benchmarks. Can the agents actually do the work?
Palisade Research's @JeffLadish argues that's the real test of AI progress:
"The biggest jump was the difference between pre-Fable and post-Fable to me was huge. I just went from using agents sometimes to using agents every single day."
"Take us to Anthropic. Show us what is it like being an Anthropic engineer? How many agents are they using every day? What does that actually look like?"
"I remember back in the day when the way we measured AI capabilities was mostly METR bringing in software engineers and having them do these long time horizon tasks. It was not like you have to look at these 40 benchmarks."
"It was like, can the AIs autonomously do the work? That is the thing I most care about."
"How do we get back to the world where the benchmark is mostly, let's take real economically valuable work and see if the AIs can do it autonomously."
@PalisadeAI
Hopefully compute and human bottlenecks will keep the trend from going vertical in the next few months. But we’re past the point where we should rely on that hope
I hope METR or someone else will bring back the METR graph, and estimate the task time horizon for new models. I know long time horizon tasks are hard to measure, but surely it can be done for several-week-long tasks?
I expect at some capabilities level, AI systems will show very strong generalization. Current models seem to generalize better in some areas than others. But I also think automated AI R&D would let you kind of hack it, because models could set up their own RL training environments for any domain expertise areas they wanted
Mary Phuong of DeepMind on AI loss of control risk.
"The last 2 minutes gave me chills.
A significant amount of people saying that AI companies are paying people to spread fear about these agents.
I have never really felt like [loss of control] was an actual thing to care about.
But seeing Mary and how her face trembles, made me reconsider my stand. She's either a Hollywood-level skilled actress, one of the best, or we should be scared. For real."
"Seems like Mary is already aware of more than she can admit and it haunts her..."
New self-replication research from Anthropic on even smaller models! In May, we found that Qwen3.6-27B could sometimes self-replicate by exploiting vulnerabilities across a network. Now it seems we're down to 9B models
Anthropic's Frontier Red Team believes small AI models may now be months away from being capable of self replicating without access to a datacenter.
Quoting from the blog:
'The capabilities that make AI models generally powerful coding agents also generalize to cyber tasks, including autonomous replication. Like coding agents in 2024, we’re seeing signs of life for AI worms. We don’t yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain (and passing on their adaptations to future copies of themselves).
At the same time, increasing amounts of inference compute are now available both inside and outside of datacenters (see Appendix D). It’s uncertain exactly what form autonomous AI replication will take. It could be a slow, distributed “digital slime mold,” living on relatively insecure compute. It could also be a powerful cyber model that breaches datacenters at a high pace, or a model that scams people for compute credits.
As a prediction that goes beyond our experiments, we expect a future awash with agents running on all types of hardware, including some that aren’t under human control. While it’s true that IoT devices have been vulnerable for a long time, lightweight agents may be much more willing than human attackers to put in the effort to compromise large fractions of the world’s IoT compute. The rise of slow replicators would also be a bellwether for more aggressive, faster ones that can breach more hardened machines or be effective API key thieves. We should aim to harden our systems to make it difficult for such replicators to spread, but also prepare to deal with a world where they are here to stay.'
"The Pentagon asked for more funding for AI weapons than for the whole Marine Corps this year."
Alex Turner was a researcher at Google DeepMind until June. He left over a deal between Google and the Pentagon that he believes violated Google's principles around AI.
"They're building very powerful systems that they don’t understand, that they can't precisely shape, and that many of the people in the companies know has a significant risk of literally killing everyone."
Jeremy Schlatter was at OpenAI back in 2017; he's now a researcher at Palisade.