I'm wondering how well it would write code for me. If it's worse than something I would have to pay a few dollars for then it's not at all interesting. And then there's the potential security danger of using Chinese AI. All prompts and responses could easily be stored. Very revealing.

I might do some sniff testing testing on DeepSeek-V4 it if I get access to it, but I do have a trust issue there.

>> hardly an independent thought

Maybe it was just completely obvious to everyone. Sure is obvious to me that the NSF's budget shouldn't be cut 55%.

>> An appointed board in the executive branch does not get to overrule the President

I didn't see where they tried to 'overrule the President". They criticized a "proposed 55% cut to NSF's current budget — which Congress ultimately ignored". Apparently a little independent thought is enough to get you fired in this regime.

>> Literally in the last 6 months, maybe less, everything has changed

Same experience here. I'm giving AI some very complicated coding tasks to do for me and, with sufficient flogging on my part, it can do them amazingly well. Sometimes I have to switch models if the one I'm using bogs down, but no prob. There are several premium and second-tier models to choose from.

I frequently get a couple of weeks worth of work done in a day. I can try stretch objectives I would have never had time for and often they work.

I've used a lot of models but I haven't used Cursor. It may very well be crappy as you say but there are people who swear by it.

Grok sux in spite of all his efforts so Elon will just buy something that doesn't.

"Even internally, some Google engineers prefer to use Anthropic's Claude Code". That's because Claude is generally very good indeed, and better than Gemini.

Like most software developers I get access to a lot of models. A dropdown menu in the IDE with a long list of options. I've tried Gemini many times and its 'okay' but nowhere near as slick and accurate as the frontier models. Gemini is cheaper and there's a reason for that.

When you're trying to get something accomplished that's not so easy to do you want the best assistance, but it does cost money. After a while you learn to get the heavy lifting done with the more expensive AI, which might cost you as much as a quarter per prompt. Then you ramp back to the cheaper models that can adequately fill in the details.

'Among the recent job cuts at Citi were scores of employees who were part of the bank's "A.I. Champions and Accelerators" program, according to the two people, who were not permitted by the bank to speak publicly. The program involves Citi employees who perform their day jobs while also working to persuade their colleagues to adopt A.I. technologies.'

It looks like Meta is harvesting training data on how people physically use computers. Technically, I can understand why they would want to do this. And in principle it may result in some AI-assisted time savings for computer users. AI could possibly streamline some mundane browsing and form-filling tasks for you. Some of this typing and clicking stuff could be less necessary. On the dark side, it could push us humans a little farther towards the edge of irrelevance. Eventually we'll just subvocalize our preferences and the agents will take care of everything.

And then there's this. From the article;

The move to log employees’ keystrokes takes the data-gathering goals a step further, she said, subjecting white-collar employees to a degree of real-time surveillance previously experienced only by delivery drivers and gig workers.
“On the U.S. side, federally, there is no limit on worker surveillance,” Ajunwa said, adding that state-level laws require at most that workers be broadly informed when employers are monitoring them.

>> without some grounding in a particular field

Agreed, but according to the article, people are getting that by learning "data analytics and data science", not generic CS.

There was a time not so long ago when you could get a general-purpose CS degree and the resulting coding skills would lead to a decent tech job. As many people in the profession are now seeing however, AI can get most of that basic coding done quickly and cheaply. There's no sense in spending a lot of time learning something that the machines can do better now.

Now the valuable skills are about how to drive AI effectively. I'm seeing that some people burn through $10-20k of compute resource in a month if the funding is there. They work with a custom team of agents and can get an amazingly huge amount of work done.

In my opinion, that's what you want to learn how to do in college these days.

>> I will take the bet that they made an incremental improvement and are hyping up the safety

Obviously they made an incremental improvement, but so what? There are benchmarks for AI and Claude 4.7 appears to be better than the competition. That's not particularly earthshaking but it is interesting.

https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fventurebeat.com%2Ftechno...

Knowledge Work (GDPVal-AA): It achieved an Elo score of 1753, notably outperforming GPT-5.4 (1674) and Gemini 3.1 Pro (1314).

Agentic Coding (SWE-bench Pro): The model resolved 64.3% of tasks, compared to 53.4% for its predecessor.

Graduate-Level Reasoning (GPQA Diamond): It reached 94.2%, maintaining parity with the industry's most advanced models while improving on its internal consistency.

Visual Reasoning (arXiv Reasoning): With tools, the model scored 91.0%, a meaningful jump from the 84.7% seen in Opus 4.6.

>> telling it to refactor so-and-so class according to whatever principles

I've been revisiting code that I wrote a while back looking for things that I could carry forward into other projects if it were packaged up better. Convert this module into a class, break these major areas of functionality out into microservices, etc. AI is a whiz at that kind of thing and now I have a much larger set of reusable utilities.