Hands-on learning is an awesome thing - especially in the age of AI!
But maybe worth reframing away from “discovering” - hierarchical planning, sub-goals, action chunking etc. are well understood - this is more of a… transcript of your personal learning journey?
Which is still cool to see. But I read through expecting there to be some kind of novel results. We probably need to categorize these kinds of posts a bit better.
Hm, I get your perspective, but completely disagree with your expectation. I'm not sure I would be entitled to expect a discovery that is novel with regard to current scientific knowledge unless I'm reading a paper. The fact that it is couched as "I discovered" and posted to a personal blog should be more than enough to convey that you might not find novel scientific discoveries within.
Fair. To clarify, I jumped in seeing this ranked #1 on HN which warped my expectations.
The root of my concern is the overall trend I'm seeing on HN: Someone spends time iterating with a coding agent exploring a field that's new to them. The agent's suggestions come from existing literature, but it rarely says "this is Search on the Replay Buffer, here's the paper." So each new concept feels like a discovery and gets written up that way. The enthusiastic tone of AI-assisted writing adds to this sensationalized effect.
So my concrete feedback would be a title like "Learnings from 600 planning experiments with LeWM on Push-T" since the current one suggests something novel about JEPA world models in general. And cite techniques they've learned along the way so others can learn from source material too.
Some years ago, I went to an Internet of Things meetup in Dogpatch, which is a formerly industrial part of San Francisco. The company had a big open space in an old industrial building. Large old windows looked out on the bay, and an endless chain and shaft arrangement allowed manually opening an upper row of those windows. There was a skylight, with the same endless chain window opening system. There were ceiling fans. And there was a modern HVAC system.
None of this was under coordinated control. So as the place filled up, it got too stuffy, but not hot. The main HVAC system did not turn on. And nobody was trying to manually operate all those windows and fans.
Now that's when you want intelligent sensing and control. Sense temp, air speed, and direction, humidity, and rain. (No, don't get them from the Internet, as Nest did.) Inside air temp, CO, CO2, and humidity. Basic motion detectors to see if anybody is present.
Now you can use all that ventilation.
As the number of people increases and CO2 levels go up, open some windows a bit on the bay side. Open the skylight windows a bit and start the ceiling fans in low, blowing upward. That will bring the CO2 down but not over-cool the room. As temperature increases, open windows more and increase fan speed. As the sun sets and incoming air becomes colder, slowly close windows to narrow opening to maintain temperature. Use fan speed to adjust CO2. When everyone leaves, and motion detectors show quiet, open windows more, bring fan speeds to high, and purge the room for a few minutes. Then close all windows, stop fans, and let temperature slowly drop to idle normal, maybe 60 degrees.
Hotels often have "smart" HVAC systems, because they have rooms which rapidly go from zero people to huge crowds. They're in a competitive business, so if customers notice the rooms seem "stuffy", it costs them revenue. Schools have the same problem, but lack the competitive motivation to solve it, so they rarely have CO/CO2/humidity sensing and a control system that can use that info properly. Offices tend to beat the problem into the ground by making the air conditioning system work too hard without using outside air when it would help.
You can buy all the controls for this, but you need an HVAC engineer to get them all working right, and there's a long tuning process. That's where AI comes in. Given an explanation of all the components and a picture of the room, and some time to experiment a bit with the effects of various settings, a system should be able to self tune to the environment.
Incidentally, the IoT presentation was mostly a guy from Samsung touting refrigerators with a tablet in the door. The tablet was not connected to the refrigerator in any useful way.
It didn't do inventory or refrigeration control. It was just a message board.
Nice! I have a similar setup. You might like the custom visualization or temperature and humidity offsets I’ve made for that air monitor. Its screen and battery cause excess heat to throw off readings and the calibration is rough from the factory, I used a cheap bluetooth govee thermo-hygrometer as reference to build a better cal curve.
- buy the cheapest hardware that gets the job done
- ignore negative reviews that are software related
- upon receipt of devices, ask claude for local-first cfw
- have fun for a few hours acting as a meat proxy
- boom, cheap local-only smart home is done
Use a SBC or nas for home assistant. This process worked for me for switches, a thermostat, an air quality monitor, a thermometer, and a window fan.
Would’ve paid ~2.5x for hardware that had the appropriate local-first capability out of the box. Instead this all fit within an existing $20 subscription to an inference provider. DS 4.1 running locally was on track too just more slowly.
If you want something that works out of the box without too much work. Home Assistant Green is also a great little device. We first ran on a NUC (more powerful than N100), but then we switched to a HA Green. Passively cooled, doesn't draw a lot of power and has been super reliable. And it's well-supported by HA for obvious reasons.
Its a do-everything solution with a do-not-break-compatibility history to its design and a lot of attention driving its price upwards. These are good things as well, so...
Examples:
- SD cards are very convenient for hobbyists "experimenting" with OS variants. They are not great for performance or reliability; a soldered on storage device is common in many other products.
- The price is more than it seems - you need a case and a power supply and an SD card.
- The case situation is bad. If you want to 'set and forget' you really want a robust case, potted. Many N100's are designed for more abusive environments.
- Most use cases don't need the HDMI, Camera, Audio, Ethernet, or all the USB ports. That is board space, purchase cost, and idle power draw that can be reduced or dropped.
- It doesn't have some modern connectors, e.g. M2, because there is no space on the board, and the PCIe lanes have gone to the connectors you probably don't use.
Many stick based computers, particularly second hand and older generations, are better for most use cases. These are typically Intel N series; but are increasingly ARM, particularly at the lowest prices.
Absolutely this. For pretty much all use cases I have, for performance, power consumption, robustness, etc. is much better the ESP option or the “big” N100. Rpi is there in the middle. For some playing or prototyping I have some from the ol good times…
RPi hardware is nice because it's a fairly-consistent, known quantity. All copies of a given board have the ~same features and bugs, so an image made for someone else's Pi 4 will run on my own Pi 4. This grants appliance-like utility that we haven't really ever had much of in the x86 PC space.
Pi hardware almost never shows up as cheap as a person might wish for. It's scarce on the used market, even when prices are high. In that space, demand outstrips supply.
And it's never been "cheap". A decade and a half ago when it first showed up, folks were already comparing the price (plus the SD card, and the appropriate power brick, and the USB cable, and the wifi adapter, and some kind of housing) to that of a used PC or laptop, and the Pi seldom came out on top.
But if we're doing new vs new, then it does alright price-wise.
And it's tiny. It's got some hackability built in, with the not-USB camera interface and the not-HDMI display support, and some GPIO that can be useful for some stuff.
It's got places where it's useful. I use them in those places. (And despite that lengthy defense, I run Home Assistant on a cheap SFF Lenovo box that came to me with RAM and an SSD for $50, delivered, in pre-used form)
I personally don’t like raspian at all, and how they change things. They try to be “compatible” with very old HW, while at the same time many SW things I’ve done, are systematically broken in each new release.
My 2ct experience.
They make industrial sd cards. I use those for many porjects that need a long life. However, this one has been running so long: Samsung EVO/EVO Plus, 32 GB, made 11/2018 used since 2019 daily. Wear has written roughly 1.5 TB in the 263 days, about 6 GB a day. So good enough for production for me!
I have used industrial cards. Are much better, yes. In a real production setup, where a stop costs actual money, or the thing is not easy to access, distance or dismantle a system, no way in hell I would use one.
On a home network with dynamic addresses: No need to worry about IP addresses; no overlaps happen, no mistakes, no spreadsheet table to manage. Use names (as long as DHCP works, then so does DNS). Renumbering the network is simple when that's useful.
It still works when the internet is down. And when the router itself breaks, we fix it or replace it and things work again. It's good to have a functional spare.
---
On a home network with static addresses: Micromanage IP addresses, keep a spreadsheet table, potentially make mistakes. Avoid using names (DHCP may be eliminated but DNS is still a failure point), making things harder to remember six months from now. Renumbering is arduous.
It still works when the router is down, but: When the router breaks, we still fix or replace it anyway. It's good to have a functional spare.
---
In terms of SRE: I don't need 5 9's of reliability for my smart light bulbs' IoT controls at home. They all still have some manner of physical switch, and they all can still be toggled on and off the old-fashioned way. Parallel controls work well-enough.
Everyone will get by almost-normally while I fix the router, which is a task that I'll set forth to be doing anyway for all of the reasons that household internet access provides.
I've recently come up with what I think is an even better solution. If someone's already done it, I have no idea. I just went ahead and implemented it because I thought it'd be interesting.
The concept is that, instead of assuming that IP and MAC addresses are at all reliable, to "fingerprint" devices on the LAN based on things such as SSH public keys, TLS SPKI fingerprints, UPnP device UUIDs, mDNS, common TCP port profiles, and IEEE OUI manufacturer in order to form mostly-stable identities. The IP and full MAC are only fallen back to when there's absolutely no other data to work off of. My software uses a daemon that scans the network in order find devices and update existing device identities if IPs change. It serves hostnames for these devices on the `.home.arpa` TLD and creates synthetic IPv6 addresses for each device in case one would rather use their own DNS server or hosts file.
I recently "finished" this project, and it's kind of awesome because I was able to swap to a new router and everything just kept working as-is without having to think about it. Of course mDNS solves a lot of these problems already, but not every device has mDNS.
What's really cool is I recently realized that this is also very useful for accessing your local network with a VPN like Wireguard. Even though I only planned on this being something you run on individual clients, I put it on my OpenWrt router and configured it with dnsmasq so that, when I connect to my network with Netbird, all of my local hostnames just work on my phone by virtue of connecting to the network and using its DNS server. That said, my project doesn't require router admin access at all. It can even run as rootless if TUN access (for synthetic addresses) is turned off.
Maybe I'll release it publicly at some point, but if anyone wants to steal this idea and do their own implementation, I am not territorial over it. It's made getting my LAN devices working together a joy.
Hypothetically, a long DHCP lease should let you ride out a short Internet/router outage. After that, hypothetically of course, APIPA addresses would be assigned, and I suppose that mDNS should be working even after APIPA addresses take over. I was unable to make this happen. I also tried IPv6-only LAN, which my brand-new printer didn't like very much.
And don't forget to backup your router and switch configurations!
i use DHCP reservations since im a bit lazy and it seems to work good. It requires a reboot of the device on the network. at least for openwrt based routers.
Presumably you're not concerned about claude deciding you're a bad actor and having your account closed? It seemed to go through a spat of doing that when people were doing the tv rooting recently to remove trackers/ads.
Like I said, DS 4.1 was on track to solve the same things while locally hosted. Using claude is a convenience. If anthropic really cracked down on RE, I’d move that workflow to open weight providers for a few dollars worth of DS and Kimi API costs, or continue to invest in my local setup.
Thread over Matter is not fully local. In Thread < 1.4, devices can connect to internet servers through NAT64. Most Thread Border Routers (AFAIR Home Assistant + OTBR is the main exception) do not allow you to disable NAT64.
In Thread >= 1.4, devices can request globally routable IPv6 addresses. The Thread 1.4 white paper does not hide the intentions:
The Thread 1.4 Border Router acts as the
IPv6 router between the Thread network and external networks,
including the user's home network and the larger internet. It is by default
configured to support external communication [...]
Cloud Connectivity: Direct internet connectivity enables Thread devices
to easily connect to cloud services. This facilitates remote monitoring,
control, and firmware updates, enhancing the user experience and
enabling remote device management capabilities. Device vendors can
now directly service and support their deployed fleet of devices,
regardless of which application ecosystem each device is integrated
with.
This is what you get when you let Google, Amazon, Apple, and Samsung design a smart home standard. They'll use every opportunity extract analytics and increase control. You'll own nothing and be happy.
Zigbee, Z-Wave, and ESPHome over WiFi/Thread are the only major modern local and privacy-preserving smart home standards.
Matter over WiFi is also not fully local for obvious reasons (though you can try to block them on the router/firewall and/or put them on a separate VLAN).
I think this interpretation is a bit too negative. Undoubtedly the companies want to spy on you, and with Thread 1.3 they were already perfectly capable of doing so - via your Thread border router, which connects to the normal LAN and can talk to the Internet (if you let it). Allocating Thread devices from global IPv6 blocks (GUAs) doesn't change that, it just removes the NAT between Thread devices and the border router.
Before Thread 1.4 the IPv6 addresses used by the Thread network were all ULAs - i.e. the devices were only routable locally and didn't have a global IPv6 address, so if you did happen to want to talk to your Thread devices over the internet, it was possible but the Thread border router had to do NAT for you.
With Thread 1.4 the border router can assign device IPs straight from the global IPv6 block assigned by your ISP, which means you can directly access your Thread device over the internet without the NAT provided by the Thread border router.
If you don't want to locally host your smart home automations (you and I do, but the average consumer doesn't) this can be very useful. It's also really helpful if you want to access your own devices remotely over the Internet and not need to bother with port forwarding.
IMO this fits perfectly with the philosophy of IPv6 - every device can have a globally routable IP address without needing NAT.
If you don't want your devices reaching out to the Internet (like I don't), the solution is to put all that stuff on its own VLAN with no Internet access (which I do).
I think this interpretation is a bit too negative.
You wrote a long message to say that you agree with me technically, but then said that it has good uses as well. I don't disagree with that, but the problem with these companies is - you give them one finger and they take your whole hand. Besides that,
If you don't want to locally host your smart home automations (you and I do, but the average consumer doesn't) this can be very useful. It's also really helpful if you want to access your own devices remotely over the Internet and not need to bother with port forwarding.
this implies that this is not possible with Zigbee. You can use your devices remotely with Zigbee, but it is fully governed by your Zigbee hub/coordinator. You can use Home Assistant, the IKEA hub, the Samsung SmartThings hub, Homey, etc. to control your Zigbee devices remotely. The very big, but relevant difference is that the devices themselves cannot connect to the mothership to do analytics or to lock away functionality. There were already some Thread + Matter devices that required talking to the manufacturer's servers to get all functionality (I think e.g. some Nuki devices, but I might misremember).
If you don't want your devices reaching out to the Internet (like I don't), the solution is to put all that stuff on its own VLAN with no Internet access (which I do).
Putting Thread + Matter devices on a different VLAN than e.g. your phone which you are commissioning it with is very fragile. So fragile that most Matter documentation just says "your phone needs to be on the same WiFi network as your smarthome hub". It is possible (I had such a setup for a while), but requires that you e.g. do allow mDNS advertisements across VLANs (since Matter relies heavily on mDNS). Since you are talking about average consumers, Zigbee will give them privacy without any extra work, whereas with Thread + Matter they have to set up their network in a way that their ISP-provided router most likely doesn't even support.
Thread + Matter has no point outside very large deployments and giving a bunch of companies the opportunity to start data mining your sensor data as well.
> There were already some Thread + Matter devices that required talking to the manufacturer's servers to get all functionality
The most I have seen that is rather the functionality not being available in matter (ie the standard does not support it) so they are only available on zigbee with the manufacturers app.
In my personal experience
> Zigbee will give them privacy without any extra work
Was exactly as fragile and dependent on manufacturer infrastructure for some fratures as you describe MoT. (the privacy part is true though, and as mentioned all the time: I have heard others had much better experiences than me with zigbee)
The most I have seen that is rather the functionality not being available in matter (ie the standard does not support it) so they are only available on zigbee with the manufacturers app.
No, it's also used for remote access, see e.g. Nuki and Tado.
Not super-nefarious, but this is probably just the beginning. If companies can gather data they will and if manufacturers can add features that require subscriptions, they will. The line always has to go up and Thread offers them the means to do so. If you give the opportunity, this always happens (see e.g. TVs, 'owning the glass').
Was exactly as fragile and dependent on manufacturer infrastructure for some fratures as you describe MoT.
I have over 50 Zigbee devices and never have a single issue (using HA and zigbee2mqtt, some on a separate Hue network). The Matter over Thread devices that I have, regularly drop off the network (no interference, etc., they just still seem very buggy). This also makes complete sense, since Thread adds a lot of complexity (IPv6, TREL, communication between border routers, etc.).
> If you don't want your devices reaching out to the Internet (like I don't), the solution is to put all that stuff on its own VLAN with no Internet access (which I do).
Or, if there is already a perfectly functioning open standard that just works and it's private by default (ZigBee) boycott the new standard created and forced upon us by BigTech and don't buy that hardware.
I tried to hop on the Matter bandwagon until I found out that it's basically impossible to keep Matter over Wifi devices on their own IoT VLAN.
I have therefore come to the conclusion that there's literally only downsides to adopting them over Zigbee.
Looks like this was published recently.
When I did research about 6 months ago there wasn't really much information other than a vague "you need to do mdns broadcast and a bunch of firewall hackery" at which point I gave up as it was not worth the effort.
Still not sure if it's worth the effort but at least there's a guide :)
> AFAIR Home Assistant + OTBR is the main exception
Luckily this is my setup.
I am not opposed to the possibility of cloud connectivity (as long as it is deactivateable) and even if I weren't I think I still prefer Matter (thread/wifi) over nearly all non-matter devices I've ever used, especially all the problems I had with Zigbee devices will make me stay far away from it. (Even though I am aware that many people had gread experiences with zigbee.)
Bunch of sensors, lights and switches (sonoff (though those were wifi, much easier but energy hungry), ikea, aqara and some more I don't remember). They were often annoying to pair, dropped from the mesh, or simply didn't like to work with my bridge (first ikea, later smart things and also tried skyconnect with HA directly).
Matter wasn't flawless, especially because I fell into the trap that was mentioned here before by trying to separate things into clans without fully understanding the needs of matter. But now that it is fixed, I had little problem with adding new devices (like not having a lamp in short range to the bridge lol) or them staying connected.
Interesting. I also have many IKEA and some Aqara Zigbee devices and they don't have any issues. Did you check if the Zigbee channel didn't overlap with 2.4GHz WiFi? (Same issues as with Thread) Did you use a multi-PAN (Zigbee + Thread) coordinator? I am pretty sure that some SmartThings hubs use multi-PAN.
The IKEA tradfri Reddit is a stream of problems that people have with their Thread devices (and people longing back for Zigbee). Part of it is firmware that is still immature, but part of it is also the complexity of the Thread + Matter stack.
My impression is that Matter over Wi-Fi is the de facto standard for new smart home devices and Matter over Thread is really only a thing for compatibility with legacy Thread-only devices that got a software update after Matter over Thread became a thing.
That’s not the case at all. Matter over thread is newer. The latest range of ikea smart home stuff is all thread.
Thread has the advantage of consuming very low power so battery powered sensors and remotes can run for years. Wifi is only used because it doesn’t require buying a thread border router.
I didn’t mean to imply Matter over Thread was older. Thread itself is older than Matter, and also older than the wave of cheap ESP32 based smart home devices.
Wi-Fi is cheaper to implement both software- and hardware-wise if you don’t need the low-power requirements of Thread, most devices don’t, and the gap will be bridged by lower-powered Wi-Fi.
The other advantage of Thread is that the devices don't need to directly communicate with a border router (I think Zigbee is also like that?) unlike Matter over WiFi, where it does need to communicate directly with an access point.
But matter is a better network. More efficient and faster. Large zigbee networks with cheap devices often result in ghost devices not responding. Matter even with the cheapest 2 dollar bulbs from AliExpress have zero issues.
Matter is not the network layer. You are either talking about Thread (which uses IEEE 802.15.4 like Zigbee) or WiFi. Matter is the application layer protocol.
We have a lot of devices and never have any issues. But yeah, if you put a lot of Tuya Zigbee devices with crappy firmware that flood the network. Similarly if you put a lot of crappy WiFi devices on your network, it's a security nightmare (wise people put them on an isolated VLAN and use WiFi client isolation).
I found a few wifi and bluetooth devices that were cheaper than or didn’t have a zigbee equivalent. Window fan and air monitor specifically.
I didn’t do an exhaustive market search, just sharing my anecdotal experience. It’s a great feeling to be able to focus on hardware capabilities and price, then will the software you envision into existence via AI accelerated RE.
i do like zigbee (now Thread). however, i want to some fancy stuff sometimes that i could never do with those devices. Such as get the weather if its available. Updated sunset/sunrise times. Those are local services i use BUT i've seen it where they reach outbound to the internet when needed.
Similar, with z-wave instead of zigbee. Except after installing two smart switches, a motion detector, and two smart outlets, I realized there aren't really any automations that I actually want. My porch light turns on and off at sunset and sunrise, through an electronic rube goldberg machine that could have just been a single photocell lightbulb.
Not contradicting or criticizing, just describing my lack of creativity.
I had a similar issue, especially for lighting, there is just very little to no automation I want. Which was a "problem" when getting into smart home stuff, because lighting was somehow the poster child and most recommended part.
My favorites are now air quality and climate measurement (the reason why i tried it again was mold prevention), climate control and a smart lock because i was scared of forgetting my keys when moving away from any friends I could hand a spare.
I pity those who've yet to experience running uncensored LLMs locally. Sure, they're not as fast as Claude, but they're still quite capable if you're patient.
It was helpful for reverse engineering some perfectly good software that the publisher recently discontinued and bricked for existing users. Sure, you could probably try to fool Claude or GPT into doing it, but I'd rather not play mind games with my LLM to get it to do something it thinks is illegal. I also just like brainstorming with uncensored Qwen because it doesn't constantly spit cautionary verbiage and disclaimers like frontier models do. The Splash version of regular old Qwen-2.6 also just did a bang-up job creating a simple daemon for managing my Govee smart outlets over BLE (without Home Assistant or other crap).
Things are moving fast now, security issues are now on the agenda with various governments, these ESP32's may be patched by Claude sooner than you think.
Probably, unless chinese companies change their business model and a secure successor to the ESP family arises. Low conviction loosely held opinion, shared bc you asked.
But why does that matter? My smart home upgrade cycle is around once every 5 years. The world is changing on a dime every 6 months nowadays. Projections beyond about a year out in the smart home market are useless to me as a consumer.
Genosys Bio - we're a London based (SF soon!) startup. We teach our models the regulatory language of DNA and use it to make synthetic 'volume controls' that let us make more of an antibody and less of a toxin.
The cool thing is since we don't change the structure/activity of an antibody it doesn't require the usual 1B + 13 years of clinical trials. Instead we cut down on the amount of times it takes and also increase production yields.
Do add me me on linkedin (Taiga Clarke) if there might be some synergy there :)
T430s as a mobile linux workstation for garage projects. For anything else, 14” M series macbook.
Thought I’d put Asahi Linux on it but the workflow I’ve settled on is to use my Mac as a thin client for a linux box under my desk. Best of both worlds.
> What if it can be Einstein, but can’t draw a Pelican, write a solid college-level essay, or fold clothes?
Then it’s an expert system.
Stephen Hawking wasn’t very good at folding clothes.
The ‘General’ part of the term ‘AGI’ seems like a trap to me, because there will always be new workflows to master. Can Astra one-shot level completion on some yet-to-be-released video game? If no, does that mean it’s not yet ‘Generally’ intelligent?
You won’t get pure ‘general’ intelligence until you find Einstein’s hidden variables and load the state of the entire universe into context.
Meanwhile, building a series of expert systems targeting specific valuable workflows is useful today and seems like it’ll continue to scale to cover huge swathes of economically valuable workflows.
I think that’s the more interesting thing to be measuring. The surface area of useful economic workflows that can be addressed with expert systems built with today’s tech.
Hitting some ‘Artificial Expert Intelligence’ coverage threshold on economically valuable workflows is what will matter for humans well before pure ‘general’ intelligence.
The only important part of 'general' is the ability to learn from experiential data and update your own model. That's what leads to general capability. Humans can't oneshot any task natively, but we can practice for a while until we uncover often novel methods of accomplishing something.
Therefore: the current transformer architecture is fundamentally incapable of AGI because the models have no mutable long-term memory.
You only have weights (large immutable memory), or context (small mutable memory).
Humans have mutable long-term memory: I can learn a new skill, adapt an old skill to new information, or learn new knowledge today that I couldn't perform/didn't know yesterday. I don't have a training cutoff.
Context engineering is an attempt to paper over this limitation. You can get really far with context engineering and huge models, but you will never get to AGI because there are many tasks where humans' mutable long-term memory outperforms.
For example, a human can invent a new musical instrument and then learn how to play the instrument they just invented. That's inference (inventing an instrument) leading to training (neuroplasticity). Humans have the ability to train our NNs with considerably fewer training samples. Everything that you can do with transformers is in one causal direction: training -> inference.
So if we take a huge with enough compute (CPUs, b200s, petabytes of SSDs), we install on it both the Astra, and the toolsuite to incorporate new sensory inputs (threads/sessions), camera, microphone, temp sensors, the lot, into a new version of the model. This model is then swapped for the old model, or traffic slowly brought over, or even adjusting weights in place.
Then my hypothesis is that thing as a whole could achieve AGI.
This feels like a very close approximation on how we humans evolve our brain. By encountering new experiences/sensations, classifying them as negative or positive to us, filling it away in neurons. Or by training motor skills etc. In the end we get more connections between neurons in our brain and we are capable of more.
Bingo, LLM architecture just does not lend itself to becoming AGI. They can get really good, sure, but they will always struggle with novel input and scenarios.
The more training data that is shoved in to them, the more they'll seem to solve novel situations, but in reality it'll be things that exist in the training data.
Aka Star Trek hologram characters aren't sentient, and actually anyone who things droids in Star Wars can think of a weirdo. C3-PO just kept running out of context and trying to revert to it's system prompt.
> Can Astra one-shot level completion on some yet-to-be-released video game? If no, does that mean it’s not yet ‘Generally’ intelligent?
If a model can't learn on their own to play some new game just as well as humans do, it's not AGI.
It's okay if they would take some hours or days of learning (like humans might), but if they can't do it at all during their normal operation, that's not general intelligence
> You won’t get pure ‘general’ intelligence until you find Einstein’s hidden variables and load the state of the entire universe into context.
But humans have general intelligence. AGI is about matching human ability, and we know this is possible in principle because brains exist
But maybe worth reframing away from “discovering” - hierarchical planning, sub-goals, action chunking etc. are well understood - this is more of a… transcript of your personal learning journey?
Which is still cool to see. But I read through expecting there to be some kind of novel results. We probably need to categorize these kinds of posts a bit better.
reply