But interestingly every now and then I look at the compaction result and it now says if you need to reference the previous conversation you can open <file>. So technically that context is connected.
I’ve noticed MCPs get unstable after compaction. but even that’s been less so lately.
Very cool work! We spend a lot of time thinking about "robust representations" in the video space. Are there any alternative ideas to JEPA right now, when it comes to speech encoding that couples meaning and sound?
For context in our domain of, video-JEPA hasn't proved to be as helpful as one would have hoped. It's decent at high level semantics (e.g. action detection) but doesn't capture enough "detail" (intentionally so) to be used as a powerful enough encoder.
I'll be sad when Google Fi is eventually killed. It's honestly amazing to have a service that's purely transactional. No notifications, no upsells, no "oops we had a data breach" (except the time it happened upstream), no roaming. Just a monthly payment exchanged for service.
That sounds high. I mean, if you paid for the 20x max plan you’d be capped at around 200/month and at least for me as a professional engineer running a few Claude’s in parallel all day, I haven’t exceeded the plans limits.
Congratulations on the launch, Edgar and Lewis! I tried out https://pg.runcaptain.com/ ---
1. I can't seem to select the text during text generation without it being deselected as soon as more text is generated.
2. It seems like it tries to emit citations, but doesn't emit proper links and instead just wrote [filename].
> one of the most common pieces of advice Y Combinator gives to startups [153_do_things_that_dont_scale.pdf].
I've been saying this for a while, but if I had to use Grok for anything programming-related I'd feel very sad and unproductive. I was playing around with a local TTS model codebase but having some issues getting it to work, so I tried explaining the problem to all the major models to see how they performed. Grok performed the worst by a significant margin, and the worst part was that it easily became stuck trying minor changes that didn't solve the key problem.
If we are to take any claims of Recursive Self Improvement seriously at all, then having a competent coding model seems like a key asset where you need to guarantee that you're remaining competitive. Why wouldn't you make coding models a top priority if you expect it to ultimately help your internal teams become more productive and effective?
There's also not an unlimited supply of researchers and engineers for them to keep burning through people at the rate at which they've been working. Although I guess for people with short timelines it makes sense to sprint hard, while people with longer timelines are more likely to treat this as a marathon. Maybe the years of burning bridges and developing such a toxic reputation are finally catching up to Elon. I think part of the harm that Elon has done is framing all the work in xAI as engineering while being highly dismissive of research, but a lot of research requires running experiments or thinking about problems and exploring them for long periods of time. If you're just grinding out work nonstop you don't really have time to let your mind wander and explore new ideas.
Honestly, I'm surprised they've done such a terrible job with programming. I remember around summer last year it was quite apparent how far behind they were with coding tools, but Elon was posting about taking that domain a bit more seriously. Why didn't any of those efforts materialize into real outputs? Something must be truly dysfunctional inside of xAI for them not to be shipping anything at all, especially considering Elon's propensity to ship undercooked products while continuing to iterate on them, as he has done in many previous cases.
I've noticed that Elon has also gone very hard on social media posting a ton of criticisms against the other big AI company CEOs like Daario Amodei. This suggests to me that he must feel very threatened, otherwise he wouldn't be resorting to such childish behavior. He must feel incredibly frustrated that no amount of money is able to make him more competitive within the AI space.
"Engram is a family of keyboard layouts optimized for comfortable and efficient touch typing in different languages created by Arno Klein. You can use the open-source software to create new key layouts optimized for different languages."
[{ 1| 2= 3~ 4+ 5< 6> 7^ 8& 9% 0* ]} /\
bB yY oO uU '( ") dD nN gG vV qQ #$ @`
hH iI aA eE ,; .: tT rR sS cC zZ
kK jJ xX wW -_ ?! mM lL fF pP
If you’re “helping a kid” then I guess I can help you. Help is criticism delivered with a constructive tone. Criticism can be helpful if you look past the tone.
I agree with your point about human level communication and treating the recipients like they matter. I generally tend to prefer communication that is more on the blunt/direct side, but if there's one thing about communication that I've learned throughout my career, it is that the people who do best are adept at communicating well with a wide variety of people with different communication styles and preferences.
The people who try to force everyone else to fit into a specific bucket of communication style, or who refuse to deviate from their own strict communication preferences no matter the audience, those are the people I see struggle to find success relative to their peers.
the district repeatedly denied enrollment after citing license plate
recognition data that it said showed her vehicle appearing overnight
at Chicago addresses
Every bit of surveillance that can be leveraged against us, will be eventually. The drive to surveil us has a lot of overlap with addiction.
I don't understand this take. Do people think engineers go in to work to turn handcranks to keep the machines running? It's actually a credit to the automation built by the engineers he fired that it kept running!
At the time I joked that like Chaos Monkey, we should have an "Elon Monkey" to "fire" arbitrary people by sending them on mandatory vacations with no connectivity to see what falls over.
First of all, I don't want to run anyone's code without proper explanation, so help me understand this.
Let's start with the verifier. The 3rd party verifier receives a bundle, not knowing what the content is, not having access to the tool used to measure, and just run a single command based on the bundle which presumably contains expected results and actual measurements, both of which can easily be tampered. What good does that solve?
I think the reason the negative luminance is potentially important for secrecy is that it means the average of the signal you’re transmitting is zero, making it indistinguishable from noise.
Also, when you hit compaction at 200k tokens, that was probably when things were just getting good. The plan was in its final stage. The context had the hard-fought nuances discovered in the final moment. Or the agent just discovered some tiny important details after a crazy 100k token deep dive or flailing death cycle.
Now you have to compact and you don’t know what will survive. And the built-in UI doesn’t give you good tools like deleting old messages to free up space.
To correct the mangling of history, there was no "list of Jews kept by the Netherlands [pre-occupation]". There were only pre-existing Dutch population registries of all people, where the personal details collected by the Dutch had included religion, not for any ill purpose.
(The Nazis subsequently compiled a list, post-occupation, but that's not what you asserted.)
Financialization of everything is so funny to me, because even I, who is extremely stupid when it comes to big money stuff, can see not having state capacity on important stuff is insane. By that, I mean hard resources, materials, THINGS.
I’ve noticed MCPs get unstable after compaction. but even that’s been less so lately.