To the Editor:
Assemblyman Gray has some valid concerns about A11560, but a moratorium on data centers is a wise and forward-looking move. There’s no doubt that AI will make changes in our …
This item is available in full to subscribers.
To continue reading, you will need to either log in to your subscriber account, or purchase a new subscription.
If you are a digital subscriber with an active, online-only subscription then you already have an account here. Just reset your password if you've not yet logged in to your account on this new site.
Otherwise, click here to view your options for subscribing.
Please log in to continue |
To the Editor:
Assemblyman Gray has some valid concerns about A11560, but a moratorium on data centers is a wise and forward-looking move. There’s no doubt that AI will make changes in our life, society, and business – but what and when are completely up in the air.
Once we get past the novelty of computers which use language effectively, several factors rear their ugly heads. At present, AI is too unreliable for serious business applications without significant supervision, it lacks the knowledge to undertake most tasks, has no judgement, and a variety of other shortcomings. But most of all, it’s way too expensive.
The vast over-valuation of AI companies is a matter of financial chicanery and fantasy hype, not of realistic business assessment. No AI company has yet been able to propose a plausible path to profitability. A
I companies are burning money at astonishing rates, and can’t charge customers anywhere near what their services really cost. Attempts to raise prices to reduce their loss margin is already causing companies to pull back on their use – and these huge data centers they’re all planning to build can never be profitable either. Their costs are excessive, and the life of their hardware is 5 years or less; they’re unlikely to even make enough to cover depreciation.
This technology is in its infancy. Companies are madly working on how to make LLMs better “understand” the real world. They’re building assemblages of subsections with unique expertise to enable LLMs to better handle the kinds of tasks businesses actually require; just as your computer has multiple CPUs working in parallel, a math coprocessor, a graphics coprocessor, short term memory and long-term memory, and a company has Marketing, Engineering, Human Resources, Operations, etc., so LLMs need specially trained sections and protocols for what kinds of tasks they handle, when to call on them, and what information to pass, and how to format it.
You wouldn’t expect a CEO to be able to service the company’s computers, or a salesman to do engineering work – the same issues apply to LLMs.
Meanwhile, engineering moves apace. AI companies use NVIDIA boards not because they’re optimum for the task, but because they’re the best available. Multiple companies are working on optimized chips and architecture; one claims its chip can produce the same output as NVIDIA’s using only 10% as much electricity; another claims they’ve developed an architecture which doesn’t need water cooling or the power-hungry refrigeration equipment of current designs.
AI companies are working on architectures to make LLMs more efficient, while ordinary companies are discovering that smaller models specially trained for specific tasks are far more effective than the big commercial models. And, of course, there’s quantum computing on the horizon.
The massive competitive data center buildout is ill-conceived and premature. Companies have already seen that simply throwing more computer power at their models has had diminishing returns. A stock market correction and realistic valuation of AI companies would rapidly dry up the fire hose of finance they depend on, and these projects would drop like flies in a roomful of Raid. There are scenarios in which in 5 years, these data centers would be abandoned or become useless or need complete reconditioning. It’s only realistic to be cautious until the industry stabilizes and its trajectory becomes clearer.
Donald A Duncan
Hannawa Falls