It’s becoming like the iPhone, once the software has access to “Claude”, everyone in the org wants it. Finance wants it for excel, marketing and design for image generation, compliance for working with documents. Sure it’s not software engineering rates, but it increases the user base beyond software developers.
Rings true in my org -- finance is a big consumer because all the software shops who do integration work with CRMs and ERPs are garbage and take forever to implement changes and via some clever prompting, you can obviate most of that work.
One regular workflow is a reconciliation we do for events that we put on — a number of costs that are expensed, a number of costs that are prepaid until the event happens, individual registration revenue that is recognized immediately and then the corporate sponsorships that are often paid in advance but their recognition is deferred until the event happens. Previously since that involved both balance sheet, income statement and CRM reporting, we relied on an integration vendor to write custom scripts to bring all the info (poorly) into our ERP. Since then, we’ve found a tool leveraging LLMs to ‘join’ those various sources and our events people generally described the report they wanted with a template in excel and it readily created that report with “export to sheets or excel” functionality.
A report that previously took ~4 hours per month for a very expensive resource now takes 30s to validate and can be run completely ad hoc by the events managers.
It doesn't really feel like an image model fits with the "theme" of Anthropic's products.
I see Claude as much more of a productivity tool than say, ChatGPT, so if they were to release an image model, I would expect it to be useful for presentations etc. But, then the images need to be more like infographics, which are difficult for an image model to get right. On the other hand, they are much easier for a strong coding model to implement, while also allowing users to make controlled edits.
It's interesting. Until quite recently image models felt like toys, and Anthropic opting out seemed like a good decision to me. I feel like the "trained on stolen data" thing feels a lot more real with image models, since they so clearly compete with the artists whose work was used - so it felt aligned with Anthropic's image as the "more ethical" of the labs.
But then Nano Banana happened, and ChatGPT Images 2.0, and now image models aren't a toy any more - you can get real, useful work done with them.
Which is a problem for Anthropic, because companies that are buying a suite of AI tools for their employees may well value the ability to create usable posters, leaflets, and most importantly presentation slides.
Anthropic don't have anything to offer there, where Google and OpenAI have two of the best models anywhere.
I've seen plenty of people at my job that have tried to spice up their research seminar slides with some chatGPT-generated images, the only time they don't suck is when someone uses the images as a joke. There's always something blatantly off, such that the output is only usable for informal seminars. On the other hand, the approach of having an LLM produce a python script, or write out a PPTX directly with placeholders and "vector" flowcharts etc, is now a very common approach for us.
Focusing on improving the inline visualization feature and Claude design, which were both recently introduced, will likely make more money and be more helpful than building an image model and competing with the very good ones Google and OpenAI have.
That data center is running on local power generation because they failed to get power run to the data center properly and essentially exploited a loop whole in planning permission that allowed them to install local power generation: https://www.loudounnow.com/news/sterling-residents-raise-ala...
It’s the only 1 out of 200 in that area, so it’s not representative of what data centers sound like. It does show how you can’t trust the operators to do what best for the local community. It does show how a functioning government works because Loudon county increased oversight and changes the rules to stop another project like that. Setup policies to manage externalities, and don’t make ignorant bans.
It seems like they should be changing the rules to outlaw that ongoing activity rather than considering it grandfathered. If a kid buys a loud car stereo and then the city passes a noise ordinance, it's not like the kid gets to keep on blasting his stereo because he already bought it.
A functioning property tax brings in a lot of revenue for the local government. Areas of the US with lots of data centers, like Loudon County, can have 35% of their budgets covered by data centers, and the worst of it is so ugly big box buildings you drive by sometimes.
Put in place sensible rules around noise, locating in pre-planned areas, and covering the cost for electrical upgrades then let the market decide how many to build. Most people appear to be getting their information from TikTok and have developed a very ignorant NIMBY attitude.
To be blunt progress does get made by listening solely to those who get short end of the stick. Japan and China have good rail in large part because the central government can simply make the globally better choice over the objections of those nearby who lose out due to noise and other factors. We don’t need to do that, we simply need to not let ignorance win and instead regulate the externalities properly, and capture value for the public through property taxes.
"let the market decide" means "let the billionaires decide" and they have. They've decided they don't care about public opinion and are more than happy to push all the infrastructure costs, pollutions and externalities onto everybody else.
Take Kevin O'Leary's DC. Massive energy tariff credits and propetty tax deductions and it uses more power than the rest of the state. So there'll need to be electrical infrastructure upgrades to get in gigawatts of electricity. You think Kevin O'Leary is paying for that? Of course not. Utah residents will be paying for that.
The blatant lies around tax breaks and subsidies are funny too. "we have to hand them out or they'll go elsewhere". No, no they won't. And if they do, who cares? Most things in life are a collection of positives and negatives. Like someone else mentioned fracking. It definitely has a lot of negatives but (IMHO) it's way more positive than data centers. I actually think AI data centers are strictly negative, meaning they have zero positives for the state and the communities affected. I honestly cannot think of a single positive that the residents of Utah will get out of Kevin O'Leary's DC.
I've never actually seen that status page before, and I'm not clear what it's measuring. My company pays for Enterprise Cloud, and we see all the same downtime as what gets posted to https://www.githubstatus.com/
No the Enterprise Cloud is just the same GitHub.com with the same shitty reliability.
However the newer Enterprise Cloud with Residency (aka on Azure) is a separate partition and has a different reliability domain (still subject to Azures bad reliability so not an entirely compelling offer). This is what you linked.
Where is the network isolation? I want to be able to be able to limit what external resources the agent can access and also inject secrets at request time so the agent does have access to them.
File system isolation is easy now, it’s not worth HN front page space for the n’th version. It’s a solved problem (and now included in Claude clCode).
Why should he put effort into measuring a tool that the author has not? The point is there are so many of these tools an objective measure that the creators of these tools can compare against each other would be better.
So a better question to ask is - Do you have any ideas for an objective way to a measure a performance of agentic coding tools? So we can truly determine what improves performance or not.
I would hope that internal to OpenAI and Anthropic they use something similar to the harness/test cases they use for training their full models to determine if changes to claude code result in better performance.
Well, if I were Microsoft and training co-pilot, I would log all the <restore checkpoint> user actions and grade the agents on that. At scale across all users, "resets per agent command" should be useful. But then again, publishing the true numbers might be embarrassing..
reply