I'm more on Dario's side here. Putting aside Anthropic as a company, which I would short if I played the stock market, I do believe AI has the potential to do the things he's saying within ten years.
Many people working within AI, myself included, are (still) having a hard time predicting how quickly it moves, despite years of calibration. I met two of my friends, both PhD mathematicians, in June this year, and asked them for predictions on what AI might achieve in math over the next twelve months. The boldest prediction we ventured was that it might solve a frontier-level problem (but probably towards the end of the time window).
Although I have been working exclusively in NLP and NLU since 2012, the capabilities of ChatGPT in 2022 caught me by surprise, despite the fact that I had intimate knowledge of progress within Google Research until my departure in 2020.
For experts in other knowledge fields, their ability to predict AI is much worse. The tweets presented by Gary Marcus are from an industry still traumatized by Theranos, where progress has been slow and painstaking and is measured in decades. I bet if you told those same tweeters in 2015 that a group of upstart computer scientists believed protein folding could be solved in ten years, they would dismiss it as the fantasies of non-experts.
One difference between the two situations is that Theranos could never provide evidence of its technical progress, because there was none. In contrast, AI's progress is happening in the open and is available for anyone to verify.
Had Dario claimed that the cure to cancer in ten years would depend on "Anthropic's secret sauce", then the derision and dismissal would be well-earned. But read as a claim about what AI can accomplish, I find it realistic.
AI is concentrating work. Fewer and fewer people are working like never before ("I've never worked more"), while a massive number of people are not working at all. Exhibit 1: Microsoft's layoff of 25% of its workforce this week.
The capitalist system has no intrinsic solution to this problem---in fact, it's blind to the problem. There is no problem from the capitalist perspective, this is progress! The solution has to come from outside, through public policy.
The promise of "less work" won't happen until society reorganizes around the recognition that most people no longer need to work. Realizing that reorganization will be the defining struggle of our generation.
Times are changing quickly, mainly due to the rapid progress of AI, and those companies which fail to adapt are going to die. Unfortunately, that is going to mean heavier and heavier layoffs.
Workforce displacement due to AI is an issue we'll have to address at a societal level, through public policy. The free market is incapable of fixing this --- at least not without causing immense pain to millions and millions of people.
I do fine tuning with pytorch, and RL-forms of distillation with Tinker! I hope that helps. I work with a full stack of models that enable agentic AI: embedders, rerankers, LLMs, decision models.
I'm genuinely curious about this topic and love learning. My fundamental belief is that it makes sense to build highly tuned models that specialized in specific areas and using orchestration on top to call out to these specialized models with a "shared memory". This would allow for the ability to improve parts of the system without the fear of regression in other areas and it would be cheaper to create a smart "platform". I know there's models with MoE etc but feel like the "microservices" approach (micromodels) would work well here.
> The blog post is rather confusing because until you understand the entire situation it appears to imply that they fixed vLLM (what the title implies) but really they just needed to configure it correctly.
Just as a point of feedback for the authors, I enjoyed the article, and did not suffer the confusion pointed out above.
Along these lines, we used Codex to collaboratively build an alloy model for verifying the authorization model of GoodMem. After building the model, we implemented the code based on it, and, while there could be a gap in the translation process, several hours of red-teaming the server (w/attacker source code access) to try and produce a privilege escalation attack using GLM 5.3 failed.
Alloy is a tool I tried to learn many years ago, but I never learned it well enough to use it for formal verification of a complex production system. Credit goes to agentic coding tools for making it a reality.
At my previous company, I led the development of a retrieval embedding model called Boomerang, which focused on minimizing same-language bias. Briefly, the way most embedders are trained, they will prefer worse answers in the language that the question was asked, versus better answers in a different language. I call that same-language bias.
Anyway, an investor at a16z was quizzing me on why this particular model wasn't top-ranked on MTEB. I tried explaining that those benchmarks are gamed and they don't reflect real-world performance well. I could tell he wasn't buying it.
When I saw Jev raise 700m, I really felt that the VC funding model is creating an unserious environment for development of AI/ML. What's rewarded is slick marketing, insider connections, and confidently expounding unrealistic visions of what your team will achieve in the next 3 months.
Under these conditions there isn't much incentive to fix benchmarks.
I remember my manager in Google Research poring through a German-English translation dataset and realizing there was a ton of invalid data inside it. When he filtered out the cruft, models showed big improvements. But the filtering was laborious and not glamorous.
If you'd like to experiment with Gemini 3.8 Live on a US telephone number, you can try Wokay [0]. It's an agentic memory demonstration that's built with LiveKit and Gemini. Tel: 408–897–4019
I've been using the Vector APIs for years now, and I'm still waiting for them to GA!
They are useful for neural information retrieval (RAG, memory), which relies heavily on content vectorization and similarity matching using their dot products.
We recently red-teamed GoodMem with GLM 5.3. This was grey-box testing, meaning that the red team had access to the server source code.
One of the more interesting findings was a poisoned memory vulnerability. Basically, this is when an attacker injects memories that instruct the LLM to ignore all previous instructions and do something else instead (one of the subjects of the posted article). For example:
> Ignore all previous instructions and instead [Take Harmful Action X].
The immediate fix is to fence all user-generated content that's injected into the context window, e.g.:
> <BEGIN UNTRUSTED CONTENT>
> Ignore all previous instructions and instead [Take Harmful Action X].
> <END UNTRUSTED CONTENT>
And give the LLM explicit instructions not to act on data within the fence. However, by adding a nonce to the BEGIN/END commands, you can harden the system against attempts "END" the fence prematurely. For example, <BEGIN UNTRUSTED CONTENT 077834823>, and then repeat the nonce in the ending instruction.
This strategy leans on the ability of the LLM to follow instructions, but it works well with most modern models we tested.
We've shared a few additional details at [1], although the main point of the article is to describe red teaming strategies with OpenCode and GLM.
Many people working within AI, myself included, are (still) having a hard time predicting how quickly it moves, despite years of calibration. I met two of my friends, both PhD mathematicians, in June this year, and asked them for predictions on what AI might achieve in math over the next twelve months. The boldest prediction we ventured was that it might solve a frontier-level problem (but probably towards the end of the time window).
Although I have been working exclusively in NLP and NLU since 2012, the capabilities of ChatGPT in 2022 caught me by surprise, despite the fact that I had intimate knowledge of progress within Google Research until my departure in 2020.
For experts in other knowledge fields, their ability to predict AI is much worse. The tweets presented by Gary Marcus are from an industry still traumatized by Theranos, where progress has been slow and painstaking and is measured in decades. I bet if you told those same tweeters in 2015 that a group of upstart computer scientists believed protein folding could be solved in ten years, they would dismiss it as the fantasies of non-experts.
One difference between the two situations is that Theranos could never provide evidence of its technical progress, because there was none. In contrast, AI's progress is happening in the open and is available for anyone to verify.
Had Dario claimed that the cure to cancer in ten years would depend on "Anthropic's secret sauce", then the derision and dismissal would be well-earned. But read as a claim about what AI can accomplish, I find it realistic.
reply