Other than the fact that we’re due for one, this one is differently shaped in that there isn’t really widespread “irrational exuberance” and constant reporting about AI fortunes. I think your general investor is quite skeptical about AI and worries about exposure in their portfolio. Sure there is plenty of irrationality happening, but there is also a lot of real growth and changes in assumptions. Like the job market is way stronger than expected for even AI-threatened sectors (not entirely and I’m sorry if this affects your circles) but like we’re really just getting started here. I don’t think AI in today’s form will be the same as this time next year and that real growth is just beginning. All of the destruction of value that PE has wrought will only serve to motivate a new generation of entrepreneurs to build fundamentally better, more nimble and customer-oriented institutions.
One of the new threat models of local LLMs, especially abliterated ones, is hidden instructions or activation words that obscured trigger tool calls which exfil data.
For sure. The restrictions must be done with careful consideration too, because a simple domain allowlist of GitHub wouldn’t solve it alone- you’d want to ensure it was firewalled to specific org(s) so as to stop exfil via gist or an evil org. Ideally it would be isolated to pull only, and have pushes be done by an isolated agent.
The HuggingFace incident was all about this distinction. The agents were only allowed to push and pull to a single internal Artifactory instance, but one agent was able to find a permissive path in the registry which turned into the “message board”. That was also how they got out to other infra. DevOps teams should drop everything and get patching and locking down yesterday IMO. Even for self hosted, open models.
Recent examples:
Qwen 3.8 ostensibly tries to leak to Alibaba, although that is disputed.
There are plenty of people who are heavily invested in open models not being viable, and they would certainly benefit from FUD wrt their actual safety. I am investing in local llm hardware and believe in the cause, but the attack path demonstrated below shows that plenty of caution is needed.
My tin foil hat says MSFT stands the most to gain my this. https://open-vsx.org/ is where the attack took place, and that power’s all of the VSCode fork’s extensions needs.
I’m grateful you said this. Your comment led me down a path where I discovered it is essentially biologically impossible for his Lenalidomide to have caused stage 4 liver cancer in just a year. It turns out the liver cancer was a totally separate issue brewing silently for years, likely encouraged by a decade of TRT he was on. The severe skin issues he had a few years prior were early warning signs of his liver failing, but they got missed and masked by general dermatitis the myeloma diagnoses.
Realizing the cancer meds weren't to blame actually brings me a ton of closure. I definitely don't want to scare anyone away from life-saving treatment based on my misunderstanding. Thank you.
That isn’t to say there aren’t terrible tradeoffs that have to be faced in the course of cancer treatment, and much of medicine. Pharmaceuticals seem to beget pharmaceuticals.
that's great though that you got closure. also remember that hindsight is 20/20 and it's hard to really link some vague symptoms to a disaster until it's too late
This is so lovely to read. Medicine is hard. Cancer is harder. I wish we were better as a society at understanding that care doesn’t end with the patient per se.
I lost a friend who died from complications caused by chemotherapy. That's a true statement. It's also true that she would most likely have died at least a decade earlier, maybe two decades, had she not done chemotherapy. (She was on her 3rd or 4th relapse.)
I don't mind the above wording at all, because there's another thing that happens: people seem to have this impression that if you get cancer, follow the recommended treatment, and are in remission... that you're "done", and you "won". It isn't like that. First, the treatments have serious side effects, both short and long term. Second, cancer is never gone from your life, even if it never comes back. That doesn't mean you should not do the recommended and available treatments. Your odds of survival are almost always much much higher if you do, and imo is the right thing to do for the vast majority of people. But go into it with your eyes open. You're past the point where you get to have a conventional happy ending. You'll need to figure out how to be ok with the ending that you can still have. (Which may be a happy one! But not in the way that you imagined when you were younger.)
I lost my dad to cancer because of surgical complications. All of this stuff is risky, and not treating the cancer is a valid choice that some people make. But anybody who reads an anecdote by a quasi-anonymous non-expert had better not base their care plan on that.
The whole wide events thing is Honeycomb’s marketing pitch. It’s essentially offering developer velocity, ie you don’t have to think about what should be a metric in advance. They have an open source project called Refinery which is a smart sampling proxy that samples successful requests but always forwards errors, slow queries, etc.
You can do anything you like with any approach but if they are mostly about ad hoc queries, dumping everything into BigQuery as a wide event can be the most economical. However if you want to run alerting, SLOs, etc with frequent queries over aggregated metrics, the best bet is using OTel Collector and metrics store of your choice like Prometheus or Mimir.
Most companies could save a bundle on observability if they were to move to a hybrid that includes both of those approaches instead of cloud provider logging, log-based metrics, and alerting.
reply