DelphiOS @Delphi_OS

This is becoming a massive problem. I have seen amongst the work product of my students a massive adoption of AI and I believe we are churning out a generation who lack foundational reasoning skills (let alone, those already in their careers who are becoming AI-dependent). While it is a great tool, we need to find ways to keep younger generations engaged and passionate about learning. What are you proposing?

A counter point from the world of risk management (this may not be true for small/solo firms). Forcing team members to take PTO and not respond to business communications can be critical to understanding key man risks in a business. If you can’t operate when someone takes 5 days off, how can you possibly hope to operate when they leave abruptly or face a family emergency or something worse. If you want a resilient business, you need to understand your dependencies, human or otherwise.

@pmddomingos Do you think the paradigm in AGI will still be making people more productive or at that point will productivity stop being measured in a human-centric way?

Yes and no. I agree that wholesale transfer of judicial decision making. (Including the role of jury as fact finder) is deeply problematic, there are ways to deploy AI in a controlled and measured way that could be positively transformative for the judicial system. It’s easy to criticize the use of AI, but the alternative is not a well funded, well staffed, and effective judicial system. Our system is fundamentally broken and is currently relying on a system of settlement and discretion (in both civil and criminal proceedings) that imposes massive downside risk onto parties in the hope that they don’t vindicate their rights. AI can help fix that and give people the forum to pursue justice. It is far from an easy an easy call, but let’s make sure we’re comparing it to the right baseline.

Unfortunately, the ability to create is not what is leading to the decline of the film industry. It is the tastemakers. For decades we had gatekeepers that kept quality high and kept us from our baser instincts. Today, we have reverted to those very same instincts which is why we see such a decline in quality. We’re feeding people empty calories and not nourishing and fulfilling experiences. Maybe AI helps, but the decline of film is more of a symptom than the disease.

Improvements in the terrible white collar tasks like PowerPoint. Models can make passable docs, but the second I try to create professional level documents, every model fails to meet the bar. I know that coding is the focus (and I admit we can do some incredible things in HTML), but a quality white collar agent would be transformative for day to day work.

So if Anthropic is blaming the corpus of data used to train the model: 1) Does that inherently mean that all alignment work an RLHF guardrails are destined to fail? If so, should we be fundamentally changing how we think about AI safety? and 2) If this is true, how is this not a five alarm fire? The corpus of training data for AI is a combination of the worst of humanity and the horrors of ASI. If that is a hidden driving force for motivating AIs, it feels like the ship has sailed . I’m trying to be hopeful, but I feel as though @ESYudkowsky is enjoying a well earned victory lap and we’re a step closer to disaster.

Anthropic @AnthropicAI

2 months ago

We started by investigating why Claude chose to blackmail. We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation. Our post-training at the time wasn’t making it worse—but it also wasn’t making it better.

318 471 5K 4.4M 2K

It’s difficult. If we’re only looking at raw intelligence you’re 100% correct. The analogy I like is that in intelligence ASI will not be a smart human, it won’t even be a human to a dog, it will be a human to a tree. On the other hand, if you believe there is dignity and value in humanity, then being truly human is the pinnacle of that. It may not be an exclusive goal, but it is part of the picture.

Disagree, while AI may only have become truly useful and transformative in the last few months, but that destination started becoming much clearer when we got GPT-4o. Claude Code and Codex are just the natural result of those breakthroughs. We may be outrunning valuations in the short term, but this level of opportunity can certainly translate to the trillions once widely distributed - likely without any new fundamental breakthrough innovations (which there is no reason to believe there won’t be any).

8/ I don’t know if Coinbase’s decision will result in a great outcome, there are so many macroeconomic headwinds that will influence Coinbase’s journey, but I credit Brian with making a bold, decisive decision and look forward to every AI lead at a financial institution and tech company being asked about flattening organizations and getting more aggressive on AI adoption - thanks for that Brian!

7/ This brings me back to @brian_armstrong. There has been a lot of discussion here what motivated Coinbase’s decisions today, and while I have no information beyond the note that he shared, I read that letter as a leader who discovered that almost all of the man-hour heavy work that plagues regulated inancial institutions can be solved with the judicious use of agents, skills, and plugins orchestrated by a well built harness and an army of MCPs that simplify and streamline access and data governance.

With the release of GPT 5.5, the limited release of Mythos, and the release of of Deepseek v4, I want to offer some on SOTA models, the realistic future of AI, and what we can likely expect to see in the enterprise AI space (I’m looking at you @coinbase & @brian_armstrong)

This is going to be the future. The market cares about outcomes and differentiability. Until we get stronger AI, the way AI companies will be able to differentiate themselves is through offering services. This also has the significant benefit of giving AI firms a unique insight into the “how” of the sectors they serve, which will invariably become unique data that they can use to differentiate products down the line. Just like in previous eras differentiated data gave you your most, today, differentiated intelligence will do that.