I really cannot comprehend the fact that a company which runs biggest ad networks on their search engine product don't do actual search and uses AI to answer queries. If you don't send trafic to a site which publishes your ads, how do you plan to sustain your main business?
Yes that is way I have added this. It could work any OpenAI compatible model or router. I have tested it with Kimi and Qwen. You can bring your own tuned model which is better if you are doing specific work.
This was inevitable since Antropic made a fortune by releasing single app with only text generation business. They did best code generator and targeted developers and enterprise users. OpenAI did only 1.5 million dollars from Sora which is obviously far from profitable. So it is logical to assign GPU time to more profitable business.
This is actually my first impression while I am reading the post. Mentions "open source" everywhere but dude how the earth it is open source without training data.
Almost no company is going to release training data because they don't want to waste time with lawsuits. That's why it doesn't happen. Until governments fix that issue, I don't even think the "it's not really open without training data!!!" argument is worth any time. It's more worth focusing on the various restrictions in the LLaMA license, or even better, questioning whether model weights can be licensed at all.
Another alternative is to use ESP32's deep sleep mode. You can tell ESP to sleep until some event occurs. There are many options for waking up the microcontroller.
The ESP32-S3 spends most of its time in deep sleep, only waking up for about 5-10 seconds every 5 minutes. The problem I ran into was that not just the ESP32-S3, every component on the board (the accelerometer, the haptic motor driver, the LiPo battery charger, the 3v3 LDO) has a some constant, minimum (quiescent) current draw. And even how your resistors are configured (pullup or pulldown) and their resistance values contributes. To diagnose it, I would need a precision measurement tool like a ($1k) Joulescope https://www.joulescope.com/products/js220-joulescope-precisi... and probably a significant amount of time.
I have also recently been through the same steep learning curve you have and the following worked for me. Reading spec sheets is fine but nothing beats measurement if its feasible. I built a custom PCB with all the power pins for all the peripherals broken out so I could put an ammeter in series with each of them individually. Then I used Nordic's inexpensive power profiler kit 2 (search for Nordic PPK2 - its under $100). Really decent specs at 100kHz sampling rate and 100nA resolution - you connect it to a PC to see the charts. I also bought my own resin 3D printer. They are so cheap these days and it helped with iterating on designs and not having to wait days for things to arrive. PS, great post, loved it.
I had this problem on a low power design. We were making an industrial temperature sensor for something that moved and we needed to run on battery.
What I ended up doing was using a second voltage regulator with an enable pin for all the accessories. When the MCU wakes up it turns power on to the accessories and waits for things to stabilize, maybe 1ms or so. Then it does what it needs to do and before going back to sleep turns all the accessories back off.
Costs you a few extra parts and a second voltage bus and the hassle of programming but it turns "small quiescent draw" into essentially zero. Maybe the regulator has a bit of leak but it should be pico amps or less.
Wait until you look at the leakage current of capacitors..! Very poorly specified, if at all, and can actually swamp the consumption of active components in these low or sub-microamp situations. The dual voltage rail that msanford described is the way to go here, gate as much as you possibly can and really focus on reducing the duty cycle.
ESP's deep sleep is not great - the datasheet for the C3 says 5 uA. That's an order of magnitude above low power microcontrollers (e.g. ATSAML), and two orders of magnitude above an ultra low power timer. Not horrendous, but higher than I'd prefer for a tiny watch battery.
Everything is out in the open nowadays. Kids can start learning whatever they what an younger and younger ages.
A perfect example is chess. It used that a lot of knowledge was in books, often in foreign languages. Nowadays everything is out there in the open and additionally you can casually play games against top 100 opposition once you are okeish enough accelerating the development even more.
The context of the log messages seems to imply that it's part of an automated (or semi-automated) string decryption system. Many binaries obfuscate strings using some simple encryption algorithm to defeat trivial analysis (i.e. "strings"), and manually reversing those is tedious. It's nice to have an automated decryption routine, but not so nice if it consists of just eval() with no sandbox...
Even if the binary isn't malicious executing random parts of a program has high risk of breaking or at least polluting the current system. Most executables will have some side effects on the system due to interactions with file system or other processes. There is no point in having executable which only crunches numbers in a loop with no sideeffects (unless it's a benchmark or something like ls executable) so in most cases there will be sideffects. And doing them in uncontrolled manner will be pure chaos. Something as seemingly harmless as backuping software can overwrite stuff if not executed in correct way.
Assuming sandbox is working correctly any parts with sideffects will probably not be very productive for deobfuscation anyway. Such automated decryption strategy would work best for purely functional parts of code. But the parts that interact with os and filesystem are either not part of decryption code (so no point executing) or if they are part of decryption they are probably the parts which try to defeat such automated decryption by detecting sandboxed/emulated environment and stopping decryption (again no point executing it without some manual intervention or very carefully designed fake environment simulating a specific target).
reply