You should seek out the recent movie “Nirvanna The Band The Show The Movie” and see it knowing as little as possible about it. Has nothing to do with the grunge band. Another Canadian comedy gem about male friendship. It might be in theaters for few days more and soon in digital format.
To be neither accelerating or regenerating? Granted the computer can make those adjustments to with much greater sensitivity than I can with my foot. Do I have this wrong?
Wait, maybe I misunderstood things. I thought in order for a permanent magnet motor to allow an axel to spin freely, the inverter would need to apply a voltage to offset the natural drag force from the rotating magnet. Isn’t the sensation of coasting the same as feathering the throttle in iPedal 1/
Yeah that makes sense too. I can’t tell if I’m imagining it or not, but I swear I’m way more aware of the background vibration/rumble from our ICE car now. Even if it’s not audible I wonder if my brain is still attuned to that sensation when it changes frequency.
I’ve seen arguments that another reason that EVs might be more likely to induce motion sickness is that they don’t have as much auditory signal that matches the acceleration compared to an ICE vehicle. There’s no equivalent engine rev sound to indicate the car is about to speed up, or not as much.
Seconded on this form factor. If you want to carry some significant load, having smaller than standard wheel lowers the center of gravity significantly and improves stability.
Having the cargo capacity also enabled us to make more unplanned stops while out and about because we could easily carry extra things that we hadn’t planned for ahead of time. Things like groceries, misc home necessities, or sports equipment.
I do end up worrying about secure parking or storage at destinations for extended stops since they are higher value targets.
Our usecase was for transporting kids se we ended up with a long tail cargo, which also feels surprisingly natural.
I also have found that an ebike lowered the effort barrier so that even more trips were attempted via bike than they otherwise would have been. I think ones that are only pedal assist without a throttle maintain the natural bike feeling.
You can also run an EV’s V2L output through a pair of AC/DC inverters outputting 8amps at 72V into the XT60 solar inputs while the batteries are providing 240V AC output to your panel.
These two batteries do 7.2kw 240V AV whole home back up, are on sale for a few more days and qualify for the 30% federal tax credit. They have the new delta pro 3s as well. They are portable and can be charged via j1772. www.costco.com/.product.400...
Readwise Reader. It’s a paid service but I think the features and integrations are worth it.
readwise.io/read
@simonwillison.net distributes a version of datasette as a standalone electron app. But he also has a version that runs locally in the browser via pyodide and wasm, no install needed. github.com/simonw/datas...
github.com/simonw/datas...
Great overview of the overlapping scopes of many tools: alpopkes.com/posts/python...
That makes me less likely to plug in at home each day and I try to wait until I get the free charging from work. Occasionally, that has backfired and an impromptu trip means I’m more stressed about my battery level than I normally would be. Also MA electricity is frustratingly expensive!
I do wish there were some locations that offered level 2.5. Something like 20-30kw to recharge over a meal or shopping trip. I can charge at home but have free 6kw charging at work 1 day per week.
There’s a fan made one man audio book project where he does voices and songs and uses the score from the movies. It’s pretty great! It’s the bluefax hobbit!
Wild! Are the weight initialized independently and/or the training data shuffled independently?
If you aren’t planning to support accessing the attributes by positional index as in a regular tuple then dataclasses should be the default. Otherwise, it could be a breaking change if you decide later to switch form a named tuple to a dataclass.
I think the merges should be added to the vocabulary so that the vocabulary grows each iteration instead of shrinking. In your example “l” and “o” would become UNK when they aren’t in in the sequence “lo”
I don’t think that is correct. Tesla Destination chargers are AC chargers and offer ~6kw typically. They are usable now with a NACS to J1772 adapter like the Teslatap mini. But the adapters Hyundai just announced are for superchargers, and don’t have the pins to allow AC charging. Only DC, via CCS
Ugh. *renders
Leveraging type hints and Google style docstrings is where it stands out. It automatically does cross linking to other type hints and redness and markdown you have in your docstrings. So you can even have example usage in your docstrings with fenced code blocks.
mkdocstrings.github.io/python/
If you want to ask questions over even more content than can fit in the prefix, then you send up storing all the content in a searchable database. Step 1 uses the question to fetch the most relevant pieces of content. Step 2 jams those into the prefix (also called context window).
This kind of a system is typically called “retrieval augmented generation” or RAG. Basically the content you are asking questions about are jammed in a hidden prefix of text before your question as input to the model. It’s typical now that multiple books of text can fit in the prefix.
Eventually, I think we will see them emerge as useful products. Alexa and Siri are more or less standardized systems and users understand their limitations. There are many different approaches right now using LLMS that feel similar on the surface but vary in behavior. It’s exciting but chaotic!
Yes! But this is the quirk of the usefulness LLMs: they are most useful to people who have the expertise to quickly judge whether the output is correct or not. They are deceptively tricky to figure out when they have failed for users that don’t already have some relevant knowledge.
Maybe it’s an obscure artist or one that didn’t exist in the world when the model was trained, and it just decides that the most likely artist in this scenario was the Beatles, with no evidence to back it up.
LLM’s are notoriously flaky at doing math on their own inputs (see the famous number r’s in strawberry benchmark). So instead of asking it to predict character offsets you prompt it to just return the name of the artist. It might do a really good job 99% of the time, but can always hallucinate.
