Get to apartment complex last night for a $5.75 ride (normally do not take these). Can see rider on map, but they aren't moving. Great. This ride had a 7min wait timer despite not being comfort. I knew this was bad. At 2 min in (5 min left) rider still hasn't moved, so I call. Straight to voicemail. Immediately ate the cancel and left.
And she still hasn't moved.
Gotta love people.
AI has a history worth exploring, and AI Research Atlas makes it accessible without a computer science degree. Follow the ideas, experiments, and breakthroughs that connect early artificial neurons and ELIZA to today’s language and image models, with clear explanations and links to the original research. Along the way, discover “The AI Toy,” a look at how Commodore 64 owners were experimenting with neural networks in 1990. Start with the story, explore the timeline, and follow your curiosity.
This is in Nashville. Banned from BNA until Oct 2 due to my own stupidity, so I turned off my ride filters except passenger rating and accepted nearly everything offered. This was over two shifts with a 6 hour break between them.
| Stat | Value | Notes |
|---|---|---|
| Miles | 468.7 | Includes miles from home to Nashville |
| Online | 10.14 Hr | |
| Booked | 9.05 Hr | 89.25% |
| 277.05 mi | 59.11% | |
| Booked/Hr | $32.60 | Down on Month |
| Booked/Mi | $1.064 | Down on Month |
| Fuel | $57.37 | $0.12 / Mile |
Yes. If the client compact-summarizes before the model sees the prompt you just sent, that turn is not run in the context you submitted it in. You paid for a 500k-window model call. You got a call against a substitute.
That is the theft: the request you submitted had a specific working set (files, tool results, ACs, the last implementation). Compact drops turns. The model then answers the same prompt without that set. You still get billed for a turn. You do not get the inference that prompt was written for.
Two things I can cite, and one I will not pretend to have measured:
On quality: it is not "80% as good." That treats tokens as uniform. They are not. Dropping the executor's remaining methods, the 10-arg constructor, the catalog list, or a failing test is not a 20% haircut. It can make the next answer wrong. "Or worse depending on what got dropped" is the accurate part.
The correct product behavior for the turn you submitted:
• Send that prompt with that history, up to the real 500k limit. • If the next turn would not fit, say so, and compact after this response, or ask you. • Do not silently replace the submitted context with a summary and then answer as if nothing was taken.
Auto-compact at 85% (status line even ambers at 80% when it has no threshold) is the client spending your remaining window on its own summarizer instead of on your request.
One thing has become clear this summer:
OpenAI, SpaceXAI, and Anthropic are all feeling the crushing weight of bureaucracy on their models and agents as each bureaucracy expands to meet the regulatory bureaucracy being foisted upon them.
Before the release of Astra, it felt like OpenAI had been avoiding the crushing idiocy destroying Grok and Claude, but Astra is just as willing to moderate out useful work instead of confirming approval as Grok and Claude. It feels like someone has decided that AI must now act incompetently to be allowed to continue being offered as a product.
Those of you who refuse to enable PIN on your rides have no right to complain about passengers asking for names, checking your tags, etc.
The PIN is the ONLY way to ensure the person getting in your car is:
These are the two fundamental MUST HAVES on every ride.
And for God's sake, enable audio recording in the app so that you are safely and legally recording every ride, even in 2-party states. If you aren't recording rides, you don't deserve to bitch about deactivations via "false claims".
There is a fire at the very back of Bell Crossing on Lynwood Cir. Residents aren't home and no pets are involved. There is a large response on scene. If you don't need to go back there, please give them space.
| Item | Amount | Share |
|---|---|---|
| Passenger payments | $4,477.31 | 100% |
| Lyft fee | $528.73 | 12% |
| Insurance / taxes / gov’t fees | $820.00 | 18% |
| Your earnings | $3,128.58 | 70% |
| Fee cap | Below 30% cap | — |
| July adjustment | None | — |
| Item | Value |
|---|---|
| Earnings | $3,275.48 |
| Tips | $338.86 (22.4% of Profit) |
| Total cost | $1,760.90 |
| Profit | $1,514.58 |
| Rides | 183 |
| Booked miles | 1,859 mi |
| Total miles (measured) | 5,565.0 mi |
| Unpaid miles | 3,706 mi |
| Booked / total miles | 33.4% |
| Booked time | 77.09 hr |
| Online time | 120.01 hr |
| Utilization (booked ÷ online) | 64.2% |
| Rate | Value | Basis |
|---|---|---|
| Earnings / booked mile | $1.58 | 1,859 mi |
| Earnings / booked hour | $38.09 | 77.09 hr |
| Profit / online hour | $12.62 | 120.01 hr |
| Profit / total mile | $0.27 | 5,565 mi |
| Cost / total mile | $0.32 | $1,760.90 ÷ 5,565 |
| Miles / online hour | 46.4 | miles include offline; time does not |
| Average | Value |
|---|---|
| Booked distance / ride | 10.16 mi |
| Total miles / ride | 30.4 mi |
| Booked time / ride | 25.3 min |
| Earnings / ride | $17.90 |
| Item | Value | Note |
|---|---|---|
| Unpaid positioning average | ~80 mi / shift | Varies with when first ride books |
| Cash cost of that 80 mi | ~$25.60 | 80 × $0.32 |
| Unpaid-mile cash cost | ~$1,172 | 3,706 × $0.316 |
| Offline travel time | Not captured | Online clock is accurate |
First is 7 days for starting a new headgasket thread. This subreddit is not a general complaint forum. If you are here to research headgaskets, then you will easily find the pinned megathread and the multiple posts reminding everyone about the megathread.
The second was permanently banned for trolling. If you aren't here to support the community, then at least refrain from trolling.
I'm trying to get GPT-5.6 Sol to create key frames of a walking humanoid. The results have been disappointing.
Does anyone care to share a working prompt?
I believe Lyft and Uber (maybe not as directly) have been using Acceptance Rate pressure to separate the good, experienced drivers and the folks who aren't as good at discerning between rides. In June there was a flood of terrible offers that drove down AR for picky drivers. The good rides were still there, and my hourly and per mile booked earnings were steady. To counter the AR decline, I became more vigilant about enabling Last Ride after most pickups. My AR climbed, but my online time dropped relative to booked time, yet actual miles climbed because when I would be offline after dropoff, I could see the flash surges, so I began chasing them to get sticky and then drive back to where the good rides were. In this way my hourly booked rose to nearly $40 by the end of July and booked per mile rose to $1.76. Both look fantastic compared to December when I was at $26 per booked hour and $1.26 per booked mile. Yet, I drove about the same miles in the same number of days. I was taking fewer riders and overall earnings were $300 less in July compared to December. I am actually making less money, yet all of my booked metrics look stellar. I've seen a lot more drivers posting booked hourly near $40 on here, but I bet they are experiencing the same phenomenon.
ViceSharp 1.2.1 is out on GitHub and submitted to winget.
ViceSharp is a managed .NET 10 Commodore 64 emulator derived from VICE (GPL-2.0-or-later). Iteration 1 is cycle-exact lockstep against VICE's x64sc. This release is the first public line that ships a full RomM-backed game library in the desktop (Avalonia) client, plus the Xbox UWP host work that went with it.
winget (package id sharpninja.ViceSharp):
winget install sharpninja.ViceSharp
Manifest PR: https://github.com/microsoft/winget-pkgs/pull/412774 (merged). CDN index can lag a bit after merge; if winget still offers 1.1.0, either wait for the index refresh or install the MSI directly:
GitHub release / MSI: https://github.com/sharpninja/vice-sharp/releases/tag/v1.2.1
winget install --id sharpninja.ViceSharp --version 1.2.1
(or download ViceSharp.msi from the release page)
NuGet packages for library consumers (including the new RomM stack) are on nuget.org under the ViceSharp.* ids for 1.2.1.
Previously you could only attach a single media file from disk. 1.2 adds a real library:
ViceSharp.RomM + ViceSharp.Library.ViewModels (also published as NuGet packages)csdb-bridge sidecar for remote/sandboxed setups)Library / Lists / CSDb views). Xbox head gets a Game Pass-style 10-foot tile grid with the same backendYou still need your own legal ROMs and media. ViceSharp ships no Commodore ROMs. Point VICESHARP_ROM_PATH at a VICE data root (or put classic VICE on PATH). See the wiki / docs/ROMs.md.
Happy to answer questions about the RomM setup path, winget install, or how the library maps media slots. Feedback welcome.
This was driving my wife's RAV4 which has higer operating expenses than my CRV.
My primary is 2018 Honda CR-V EX. Purchase price was $30,000 OTD with 51k miles. Three months later we bought my wife a 2019 RAV4 Adventure, which was about $40,000 OTD with 105k miles. The CRV has 159k, 108k, mostly rideshare. The CR-V has been in two accidents. The actual cash value is $6500. It has depreciated $23,500. The RAV4 has 133k now. It looks identical to when we got it. It's ACV is $19,500. It has depreciated $20,500, which is $3,000 less than the CRV!
Now, here's the paradox. I was without the CRV for all of July, so had to drive the RAV4. The CRV has depreciated at $0.22 per mile. The RAV4 has depreciated at __*$0.74*__ per mile. I put 4400 miles on it in July, so my ACV depreciation was $3,221, or more than the depreciation of the CRV for the same miles.
The takeaway: your rideshare vehicle is an investment, and expense. Using you personal vehicle, especially short term, ultimately costs more to operate than you are likely to earn.
winget install sharpninja.ViceSharp --version 1.1.0 # ViceSharp
Everyone wringing their hands over Honda's EV strategy can stop. Even in California, Hybrids are outselling EVs. This was the path Honda identified a decade ago and then was forced to lose $18B on EVs by government mandates. Given how quickly that unraveled, Honda was completely correct from day 1.