DickensZhao

Dad of a curious 5yo. 10+ yrs in embedded systems.
Sharing notes on AI hardware & edge intelligence.
Building practical AI gadgets and learning tools for kids.

[DickensZhao @DickensZhao123]

Apple reportedly ships its first touchscreen Mac on Wednesday, 20 years after saying no.

I spent about 10 years writing touch firmware and algo. The bug list I'd worry about: palms on the keyboard deck, and display noise from an OLED panel sitting right under the sensor.

Source: https://t.co/2h6QbQ4rKy

[DickensZhao @DickensZhao123]

Per Gurman, touch is an extra input next to the keyboard and trackpad.
macOS will show different controls depending on whether you've been touching the screen or using the trackpad.

Rumors also mention a reinforced hinge so the screen doesn't wobble when you tap it.
That one I believe. A wobbling panel turns a clean tap into a smeared touch.

[DickensZhao @DickensZhao123]

Apple reportedly ships its first touchscreen Mac on Wednesday, 20 years after saying no.

I spent about 10 years writing touch firmware and algo. The bug list I'd worry about: palms on the keyboard deck, and display noise from an OLED panel sitting right under the sensor.

Source: https://t.co/2h6QbQ4rKy

[DickensZhao @DickensZhao123]

@InkandKeyApp Exactly.

[DickensZhao @DickensZhao123]

𝗧𝗵𝗲 𝗺𝗼𝘀𝘁 𝗶𝗻𝘁𝗲𝗿𝗲𝘀𝘁𝗶𝗻𝗴 𝘁𝗵𝗶𝗻𝗴 𝗮𝗯𝗼𝘂𝘁 𝗨𝗻𝗱𝗲𝗿𝗱𝗼𝗴 𝗺𝗮𝘆 𝗻𝗼𝘁 𝗯𝗲 𝗽𝗿𝗶𝘃𝗮𝗰𝘆.

It may be the economics of AI.

Most personal AI products today follow the same architecture:

device → cloud API → frontier model → answer.

Conway is taking a harder route with Underdog.

The models run on your computer.

Email, calendar, meeting notes, dictation and personal context can stay local, while the AI itself runs on hardware you already own.

But Conway isn't just building an app.

It's building three layers together:

models
an inference engine
a consumer agent

Its Saluki model is a good example.

It's based on Qwen3.8-27B, compressed to about 7.89GB, with a focus on preserving tool calling so a 27B-class model can actually fit on consumer hardware.

The interesting implication isn't simply:

"local AI is more private."

It's this:

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝘄𝗵𝗲𝗻 𝘁𝗵𝗲 𝘂𝘀𝗲𝗿 𝗽𝗮𝘆𝘀 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲?

With cloud AI, more usage means more tokens, more GPUs, and more inference cost.

With local AI, the user already owns the compute, memory, and electricity.

That matters a lot if agents move from answering a few questions a day to doing hundreds of background tasks.

Another thing Conway gets right: it doesn't seem obsessed with owning every foundation model.

Saluki starts from Qwen.

The differentiation is in compression, tool use, inference, memory, and the product experience around it.

That's a much more realistic strategy for a small team than trying to train another frontier model from scratch.

There are real constraints.

A high-end Apple Silicon machine is not the average consumer device. Local models still compete with much larger cloud systems. And privacy alone may not convince people to trade storage, battery life, or capability.

So Underdog still has something important to prove:

Can local AI become not only more private, but faster, cheaper, and genuinely more personal?

If it can, personal AI may stop looking like another SaaS category.

It may start looking more like a new layer of the operating system.

[DickensZhao @DickensZhao123]

@bzogrammer My son used to watch AI-generated short videos and got bored the second they stopped. That's why I'm building a parent-guided app for real-world Montessori-style play, ages 3 to 6. And I think e-ink's slow refresh could help kids here.

[DickensZhao @DickensZhao123]

@MilkRoadAI There is still a long way to go between boosting enterprise efficiency with AI and monetizing those productivity gains

[DickensZhao @DickensZhao123]

Plaud users praise summaries but also report extra labeling work. Both can be true.

As an embedded developer, I'd test AI recorders on the same meeting before and after an update, counting steps to produce usable notes. Measure the whole task.

[DickensZhao @DickensZhao123]

Plaud can separate speakers without remembering your colleagues.That makes sense from an engineering perspective, but it adds friction for users.

Its Oct 20 docs say it recognizes your own voice; other names must be added per recording. Users recording recurring meetings report having to label the same colleagues again.

[DickensZhao @DickensZhao123]

For school software, I'd test an AI-off mode before delivery: do lessons, assignments and learning records still work? Tell schools which features they can control.

Would you adopt teaching software that becomes unusable when its AI is disabled?

[DickensZhao @DickensZhao123]

NYC's school AI restrictions offer a product lesson: can your teaching software work with AI switched off?

Chalkbeat reported Oct. 23 that Magma Math disabled its AI hints and won clearance to keep operating in city schools.

[DickensZhao @DickensZhao123]

NYC's list of 38 flagged education products is disputed. Some vendors say their tools don't use the restricted AI.

Officials say restrictions can apply to features, rather than whole products. Being listed alone doesn't establish that a product harms learning.

[DickensZhao @DickensZhao123]

I'm building a parent-guided app for real-world Montessori-style play with kids 3 to 6, and this is the tension I keep running into. An AI that suggests the next activity also narrows what the kid would have picked alone. I think real world interactions should be done at least before asking AI.

[DickensZhao @DickensZhao123]

@freyadnd The raw signal is always messier than the clean number on screen. A sleep score that drops when you feel fine is often the algorithm reading noise as a trend. Showing less precision would probably help people more.

[DickensZhao @DickensZhao123]

@Tristan0x @log4j Exactly. Once HDMI input reaches the TV's scaler, it's just another frame to the SoC, which is why ACR can fingerprint your laptop the same way it fingerprints Netflix.

[DickensZhao @DickensZhao123]

Patching Brotli isn't finished when the dependency file changes.

After 10+ years in embedded software and algorithms, I'd ask: what version is actually running?

[DickensZhao @DickensZhao123]

Never wrote a decompressor? Check your Brotli dependencies anyway.

Some HTTP clients automatically decode responses when Brotli support is installed. The flaw affects Node.js and Python packages too.

Searching only your own code can miss the decoder you run.
#SoftwareSecurity

[DickensZhao @DickensZhao123]

Using HTTP doesn't automatically expose you to the Brotli flaw.

A vulnerable decoder must process malicious data. GitHub lists Python fixes in Brotli 1.2.1 and brotlicffi 1.2.1.0. Node.js updates are available too.

Check versions and actual decoding paths.

[DickensZhao @DickensZhao123]

For Verizon's delayed iPhone Duo buyers, the next promise matters too.

MacRumors showed one delivery estimate moving from October 23 to November 7. Not all orders were delayed.

What compensation and follow-through would persuade you to keep waiting?

[DickensZhao @DickensZhao123]

Would $100 make you forgive a late iPhone Duo?

MacRumors reports that Verizon has promised some customers with delayed preorders a $100 Verizon gift card, usable toward bills or merchandise.

The compensation doesn't settle when the phone will arrive.
#ConsumerTech

[DickensZhao @DickensZhao123]

Verizon's iPhone Duo delay cost one Reddit poster a vacation day.

They said they'd taken time off to sign for the phone, then received a delay notice and a $100 gift-card promise.

Earlier notice could have let them make other plans.

[DickensZhao @DickensZhao123]

@rachelgoodlad @Alphaschool The assistant-guide role gives her a real way to be useful. Have you noticed that confidence carrying over to activities where there are no points to earn?