You just need to keep an eye out for jokers like this
More funny stuff from Claude
What does a geek (yours truly…) do when he files his exact same typical ZFR OCAS departure to eastern france, eurocontrol validated but get a “Unable to activate your flight plan due to non standard routing” from London ATC and have to continue VFR?
he ask claude: "Go through all my G1000 log files in this directory, find all the flight between EGTF and the point BILGO, I either file airways (west and south departure early climb) or OCAS (east and low departure later climb). Cluster my track and tell me how much I save by going OCAS.
The result:
The details:
Ah, but is it correct or just confident jibberish accompanied by graphs?
it is correct
to be fair, there is a little bit more behind than just the prompt, as I have build a quite extensive library of parsing logs, routing, waypoint, plotting routes etc and that’s extensively tested. so really here, this is claude code building an adhoc python script using all these tools, he just had to put together the analysis, not work out all the underlying library… and the output is from a deterministic script (AI generated): numbers are from real calculation
Where did it get the controlled airspace database from?
It does not have airspace information.
What it did is run a clustering algorithm in the paths from one point to the other, came up with 3 groups which he name based on navpoint that distinguished them. But the cluster were pretty in line with the reality, so I gave him the name for the clusters… ok, that was a bit cheating but better than getting the original name he came up with: cluster 1, cluster 2, etc…
Claude has a funny SOH sometimes:
Anyway, I am going to take on Honeywell 
The code was generated by the £150/year Claude (Opus 4.8 in High mode). It is about 500 lines of C, done entirely with graphics primitives, single floats (the CPU does basic floats in 1 cycle i.e. 7ns so there is no point in integer maths; precomputing trig funcs is still worth doing since a sinf() is 1.8us). And it does get through the money if you upload screenshots of the LCD… I had to do a £35 topup yesterday. It took a cumulative time of maybe 5hrs. If I had written it, it would have taken me a week or two full-time.
I am still in the normal UI, not Code, because it suits the way I work, getting it to just write specific isolated functions
So, like everybody (who knows) says, it is a great productivity tool.
My free copy of Claude has now got jealous and is cutting me off for half a day after every interaction. I guess they are trying to make money before the IPO. Will the stock be worth buying? I doubt they can moneytise it to the stratospheric levels they are pushing.
I doubt they can moneytise it to the stratospheric levels they are pushing.
Probably, but definitely Anthropic would have better shot than spacex!!!
Last week Microsoft changed their pricing on GitHub copilot, from a flat fee to token usage base, which resulted in a huge cost surge for clients, despite reasssurance from Microsoft sales that it wouldn’t…
I think it is interesting because this highlighted the fact that Microsoft harness is quite inefficient at spending a lot of token for tasks compared to other like Claude code and cursor… quite a few people feels like they are being charged by Microsoft for their own inefficiencies and looking at switching to Claude or cursor… of course I am sure Microsoft will improve their tool efficiency but at least momentarily yet another positive boost to Anthropic…
precomputing trig funcs is still worth doing since a sinf() is 1.8us)
For the degree of precision you need here, you can replace all trig etc functions using cubic splines. Just precompute the spine and then it’s half a dozen FP operations to calculate for a given input;. You can get a circle accurate to 1% using just four points!
