By transaction volume yes. Presumably it’s not all spending though, if you regularly win back some of that money. The same dollar is probably wagered multiple times before it is lost.
Ontario lotteries pay out about 58 cents per dollar in, so that's still $13,000 annually out-of-pocket for a total spend of $31,000. That's crazy money.
Sportsbooks typically pay back more like 90%, brick and mortar casinos 90%-99% depending on the game. I don't know what Ontario mandates for regulated online casinos, but I'd guess they're at the high end of that and focus on volume, return custom and dark patterns rather than extreme house edge.
If it's 98%, the average household loses 0.2% of their income gambling.
What I want (which I don’t think exists?) is a way to trigger turns that the user can monitor in the codex application. I.e., when event X happens, my application triggers Codex to take a turn with input Y, which the user can monitor through codex. Right now the only way to get close to this is with polling or essentially rewriting a codex-like frontend.
You can give the agent a tool (or bash script) which waits for events. Agent calls it and the tool sleeps until an event happens then returns it to the agent.
I just wrote my own VR harness in a weekend with Astra. It mentioned an SDK for exactly this in passing, but it was an experimental personal project so I didn’t bother to review the code.
I was doing exactly what you’re describing. I think this is a ToS violation for anything other than personal use though.
do you mean like a cloud agent provider? What we're working on at noriagentic.com may be relevant -- you can fire events from slack/web/cli to kick off an agent in a box and talk to it as if it was running local
Two notes - 1, this applies only to a pair of base models in the Korean market (for now) and 2, the info is still there, but it's on the central screen and not on the instrument cluster behind the wheel.
As a driver I would be so freaking confused staring at the empty space behind the wheel. Maybe it is just my Gen X experience of learning to drive on a stick and paying attention to tachometer and coolant temp.
I drove a Toyota Yaris for about 10 years and those cars had the instrument cluster in the middle of the dash. It didn’t take long to get used to and one of the unexpected benefits was how much better forward visibility was. The Yaris wasn’t a big car but honestly the drivers position felt a lot more open and less cramped with that layout, and realistically the center of the dash is almost always empty space anyway. I personally liked it a lot. Not sure the “no instrument cluster + massive iPad in the center” is going to be quite the same feel or work as well, but everything being in the center feels way less weird than you would think.
> Maybe it is just my Gen X experience of learning to drive on a stick and paying attention to tachometer and coolant temp.
You and your fancy cars... my 1981 Vanagon doesn't have fancy stuff like a tach or a trip meter or a temperature gauge. Who needs a tach when you have ears, and anyway there's a colored dot on the speedometer at the top end of first, two dots at the top end of 2nd and three dots at the top end of 3rd. If you go over, the rev limiter will let you know.
While working on the speedometer, I did drive it a bit with no cluster installed, and it was weird, but I could get used to a bit more visibility, although the extra visibility was obstructed by the steering wheel.
I bet you had a separate indicator for left turn signal and right turn signal, but do you really need that? --- a single indicator if any turn signal is active is plenty.
I've made the switch before. At first yes of course you do stare at the dash in front of the wheel for a sec before remembering it's to the side but I didn't find the learning curve too difficult. After a month or so it's second nature.
I still don't understand how these massive touchscreen displays are allowed in vehicles.
I had to rent a van that had everything "integrated" in its display. Absolutely atrocious software, latency and responsiveness for standard things such as temp and fan speed.
Also withholding this "extra" means you have to take your eyes off the road and look down to the right/left to make sure you're doing the right speed. Sounds a bit over the top, but I wonder if these designers think about the fact they might actually be the ones responsible for people dying in crashes.
Kimi K3 was an interesting model only a month ago, and now we're looking at the same performance for 1/20th of the price. Wild how fast this is advancing.
Real question: is there anybody that is both maintaining alpha-dev capability by keeping abreast of all these daily changes, while also reserving enough time to actually work?
Seems like we've reached the event horizon of whether AI advances are worth paying attention to.
I feel like it's most useful to get in a bit of a groove with one setup and then poke your head up every now and again and try updating. Staying at the bleeding edge of anything can be a bit of a treadmill.
I think the play now is to just try out whatever the best new model is every time you see a headline that fundamentally reorganizes your conception of what's possible.
I enjoy using opencode go to play around with a lot of different models. I wind up using deepseek v4 flash for most everything, stepping up to minimax m3 if that doesn't cut it, finally preferring GLM for complex tasks or important planning I want to go right the first time
I recommend opencode or something akin to it to play with models. Any big model updates or hot new ones will naturally run across your desk that way
I don't think you need to be keeping abreast of them really, you just need to be using the best model you can get enough tokens from, which for many people is Fable 5 @ $200ish, ideally fanning out implementation to cheaper models
I would love to be shilling for Anthropic, but I am not. I am part of a group of about 30 developers, and 80% of them are using Fable 5 and very sold on it, with the remainder being committed to Sol. Both are competent, but among our set (who will try anything), Fable 5 is definitely winning. The fucking refusals for security work are insane though, and I hate them.
I use Sol and Grok 4.5 as my inline debuggers/reviewers, and both do well, and are decent at token save. DeepSeek V4 Flash 0731 found some interesting bugs when I tried it a few days ago, and I'm curious to see if that also joins the code-review line up
In my use, DeepSeek v4 Flash (which replaced the quite excellent MiniMax M3) lags behind GLM 5.2 & Muse Spark 1.2 (let alone Kimi K3). Also, K3 is a much bigger multi-modal model, while Flash is text-only and likely optimised for coding tasks.
Yep, and the v4 flash final is about 2.5x slower than preview making it no longer a fast model, in fact slower than Luna and bigger models in many cases.
Spark is actually the interesting one imo. It's significantly better, also significantly faster. If you are ok with letting Meta soak up your data (which DS does too) it's also the same price.
I've been using this DeepSeek model the whole day today after building with 5.6 Luna extensively over the last week and I would disagree, at least for Rust + OpenGL.
DeepSeek just spend almost 2 hours trying to figure out why terrain textures were not working. It tried everything over and over again, it even had reference code for meshes on how to setup the rendering with materials, and it could just not do it.
I finally gave up and gave it to GPT-5.6 Luna instead, and figure out in a single prompt after 20 seconds, that the terrain mesh was being initialized with None in the material slot.
Other tasks it has managed to figure out at least, but it is significantly slower than GPT-5.6 Luna and it requires a lot more iterations.
That's roughly my experience. Luna is extremely efficient and at higher levels of reasoning and longer running tasks more capable.
Reading the DS reasoning is wild, it's constantly going in circles. The most minor lack of clarity in your prompt and it will spend ages going back and forth on what you meant. It reasons 5x longer than the preview which makes it really slow now as well. We did a lot of work to nudge it to be decisive and improve our evaluation setup to there's more clarity, and it helped but only marginally.
Ours is a full-stack app one shot test so it includes backend, frontend, design, and QA/testing. It's graded by Opus xhigh and Sol xhigh and the grades are averaged.
DS4 preview would finish in 20 minutes flat on high reasoning and grades 6/10. Luna high gets 9/10 in about 30 minutes. DS4-final is crazy - at high thinking it's taking over an hour and getting ~8 but only had one successful run as I got tired of waiting so long after many early abort/retries trying to debug why thinking was so long. The lowest thinking still takes over 45 minutes, and with thinking off it actually is finally closer to preview in time but actually get's a much more varying result anywhere from incomplete to 6 it seems.
Costs per run DS4 is best but not actually by a lot as it's spending 10x the tokens with all the reasoning and mistakes. It's a very brute force model and I really preferred preview in many ways for how predictably fast it was.
Side note, Spark 1.2 is a nice model for this test, best in frontend design and fastest to get results together, though not nearly as efficient as Luna. Grok scores similarly to Spark but at like $50/run vs the contributor Spark costing $1.50.
Edit 2: btw it tests a team of agents working together in a special harness and stack. So 20 minutes is for 8 agents essentially. That said everything was built around DS as it was the cheapest to iterate against so even with that advantage the new one struggles.
Not for long, Deepseek is saying they will have a significant price jump soon. They really shouldn’t do it because they are on the cusp of capturing the scalable API market.
They need to be able to serve their market. The price increase is partly load shedding. If they improve their ability to serve their load, they can always drop it again, as OpenAI did with Luna recently.
My read is, OpenAI is neither able to claw b2b money (away from Ant) nor are they able to stave off open weights on the other. In short, they're struggling to hold onto their distant #2 position in the coding market, and these pricing changes reflect a (desperate) change in strategy.
I am also skeptical this is true. I made few typical exam essay questions and inserted the requisite clause about adding nonsense about Madagascar, and then had chatGPT answer the questions. In almost every case the output included a little executive summary at the bottom below the essay content, clearly calling out the nonsense it had added. I could believe a few students missing this, but over 90% of the class?
Generally, all long guns that are not auto/semi-auto must be cycled after every shot in order to chamber another round. This includes pump-action, bolt-action, and lever-action rifles/shotguns. There are exceptions, such as revolver rifles.
There are also instances when semi-auto pistols might need to be cycled after every round.
Surprisingly, Coriolis at extreme ranges (like 5000 feet) can be relevant enough to be corrected for, but how much and in what direction depends which way you're shooting and where on the planet you are. There's a fun calculator here: https://codingace.net/physics/coriolis_effect_shooting.html
Enterprise customers don’t get those plans, at the enterprise level you have to pay by the API rate… so people don’t have limited use, but you’re also not getting the heavily discounted rate the “normal” plans are at.
>Meta plans to spend up to $135 billion on AI infrastructure through 2026 and commits $600 billion to data center buildouts through 2028
And they can't afford a few extra billion that their engineers can utilize right now?
Looks like AI as it develops is intended to be too expensive for regular people in the long run, but if Meta can't even afford it at that rate, who can?
I’d disagree wrt “modify”. There are all sorts of tools for modifying LLM weights (ie to remove refusals, remove layers or experts, merge models, finetune, and more) and a quick glance at huggingface or civit will show those in very active use.
I don’t think the hardware requirements are relevant. If a research lab publishes the code their particle collider runs under the GPL, that doesn’t make it not OSS even though they’re the only ones on the planet with the hardware to run it.
You can also edit binary distributions of models with means besides changing their weights. See "LLM Neuroanatomy: How I Topped the LLM Leaderboard Without Changing a Single Weight."
On the spectrum of:
careful engineering--hacking--mad science
This kind of thing falls far towards the mad science end of the scale, but has proven effective.
I feel like I have a different $20 plan than everyone else. I have no problem hitting my 5 hour and weekly limits. Don’t get me wrong, it’s a great deal compared to API pricing, but it’s a far cry from “unlimited”.
I get about 20 minutes of work from my 5h limit with the $20 plan. It wouldn't bother me as much if codex would continue after the token bucket refills instead of waiting for me to show up and tell it to continue. I don't jump to the $100 plan because I would be in the exact same situation.
Harness matters in this. Using the Codex sub with Hermes eats tokens like nothing. Using it with Pi is much less but you don’t get the long term memory. When you were able to use the Claude subscription with Pi, I barely hit the 5hr limit. When they stopped allowing that, CC harness just chews thru tokens.
Interesting. I'm mostly using Claude, so perhaps I'm not nearing the limits, but I do use Codex (for coding and reviews occasionally) and use chatgpt for second opinion many times, including "pro" research. Never got to my limits. But again, not my main go to tool.
reply