CyclicalObsessive

Forget I mentioned it - sorry - need to only think about computers or my head explodes, my heart races, my palms sweat, and I need a nap.

I drive by the ultimate irony everyday in my neighborhood- a green cybertruck.

Red and yellow have developed bad reputations

… LLMs must not be “trusted” for out of domain, and must be warily reviewed when asked about edge cases of the domain. (Knowing we don’t know an answer is one of humans greatest strengths )

Exactly the issue. If carefully domain curated training is used in the mainstream of that domain, by a person also trained in the domain to enhance the human’s efforts, with the human’s critical review of the LLM output, the tool is a force multiplier.
…

> "With smaller training data sets, training itself becomes more feasible."

Used to design grammars / language models for speech recognition. Grammars: far better in constrained domains. Small language models could not match the needed reliability. Smaller needs much higher quality training.

Therapy is big ATM. Just get your button in therapy, they will teach it strategies to feel ok.

> My theory of mind is that all the other minds are brilliantly corrupted. In fact they make me feel stupid.

Paper that might provide some background to the #OpenWorldObjectDomain (recognizing unknowns)

openaccess.thecvf.com/content/CVPR...

Looking a long way down the road, if you apply for a software patent, I would like to get a mention somehow.

Or perhaps if there is a good open source RDF world object DB that could be brought local, the email could ask “is this an xyz object, if not what is it?”

The next step, “transfer training” the model for each new unknown has two easy solutions: 1) send the image to Google cloud vision (totally not privacy), or a “Mommy/Daddy, what is this?” Email with a daily question limit.

A pattern for this problem is the RTAB-map visual pattern recognition system which discovers artifacts in a pose stamped image, and compares that to known sequences of previously “seen” pose stamped images.

A language model trained to only give two answers I know that because…, or I don’t know that with option of “but I do know the unknown has aspects of these things I do know”?

“Language” models currently hallucinate at their edges of knowledge, rather than “knowing” they are at, near that edge, or beyond it. There is the open set real world recognition problem, but knowing that it does not know is slightly different.

Survival of the corrupt-est - why didn't they teach me that in school? Everyone was lying to me - "be a good person", "be fair", "care about others", "crime doesn't pay"...

I was such a sucker, believing that "you can't fool all of the people all the time" was stronger than "nice guys finish last"

In a time of “Who can you trust?”, today we learn that Trump’s yes-folk lied about the employment status 208% last year. This deception feeds the stock market fallacy, with massive CEO compensation and shareholder “return on investments”, inflating the prices of all the goods and services.

Three mobile hobby robots shown, two have Minion characters on them, one has the WALL-E robot character in it. Robot with Minion Carl is 8 years old. The other two robots have stereo depth cameras for visual simultaneous location and mapping with neural net processing for object recognition.

This has been my dream for many years, to have my robot wander my house (using vSLAM) taking pictures of “unknown interesting object”, (<-hard part), segmenting and learning the object, repeat till nothing unknown found.
#OpenSetObjectRecognition #KnowingDontKnow
(All local processing for privacy)

Can’t read it on iPad mini, too small.

I follow @msfsofficial.bsky.social and read posts with #msfs tag.

Purposes:
- (Hope to) Hear what is new with MSFS2024 releases
- Waiting to hear MSFS2024 fixed obelisk trees
- Hoping for #IndiaFoxtEcho #F35 Lightning II half-off sale
- Watching for MSFS2024 #LongEZ and #KLNA pilots

While I know four languages, Spanish is not in my brain. I thought the show was wonderful, and the music invigorating, but “inclusive” the lyrics were not. Surely subtitles crossed someone’s mind?

There is no "the People" - only DJT, DOJ, DHS now. It was not a bloodless coup, but it was a coup and "the People" don't have strong enough tools to get the horses back in the barn.

Just pissin up wind.

But then when I moved across the room to my other laptop, the same website wanted me to use my phone to scan some QR code.. It wasn’t enough to let Google make up an unintelligible string password, now they are resorting to unintelligible pictures?

My Cat Kept Jumping on the Counter..So I Built a Robot YouTube video by The Lab-X

Loshi better mind his manners

www.youtube.com/watch?v=b9DP...

Oh wow - sorry to hear of your implied loss.

I should add that the Create3 has the most accurate dead-reckoning I've ever seen with wide wheelbase, non-slip wheels, great encoders, IMU, optical flow sensor. It also has great bumper and wall ranging detectors. AND HAD the most wonderfully present support team of any commercial robot.

Amazon Astro - It was expensive, but really really powerful using both edge and cloud processing. The Astro showed me that my dreams for a single board computer home robot were not possible.

My guess is that $1500 was too expensive for a home robot with only security as the primary function.

The Create3 ROS processor was severely over worked, running at 50% at idle with many publishers turned off, and totally swamped by third party ROS topic traffic. I pushed iRobot to create a segmented DDS "republisher" to isolate the Create3, but some intrinsic action calls could still swamp it.

No, I missed out. When they recalled all the business Astro robots, I asked the team to add a ROS 2 interface to the platform. My thinking was that the ROS community needed a new standard robot platform where the TurtleBot4/Create3 failed.

nickbearded (@nickbearded@mastodon.social) Attached: 3 images Built a custom #Buildroot system for #RaspberryPi 4 running the brand new #Linux 6.19.0-rc1-v8 with PREEMPT_RT! 🚀 Real-world stress test: 1M cycles with cyclictest 2x stress-ng ...

Also a real time preempt OS kernel (FreeRTOS) for reliable sensor timing, in addition to using a non-RTOS (Fire OS) for applications not requiring frequent attention.

Here is a guy building RT_PREEMPT Linux images for Raspberry PI:
mastodon.social/@nickbearded...