WWDC 2026

[NotEntirelyConfused]

I also asked it to write me a LISP program to multiply two numbers using recursion and without using the multiply operator. (This was one of the questions on my written doctoral qualification exam in 1992). I believe its answer was correct, but i haven’t looked at LISP again since that exam, so who knows.

Then i asked it the question I often asked job applicants at AMD, to see if I could rattle them.

It’s first answer was the naive, but wrong, answer that they (used to?) teach undergrads. I pointed out the effect that it hadn’t considered, and then it got the answer correct.

Unlike the rest of the examples you provided, this would seem to speak far more to how good Google's model is than anything Apple has done. But your earlier examples are very encouraging! And regardless of where the model came from, it's good to have a capable one as Siri's brain.

 

[NotEntirelyConfused]

BTW I've encountered a really bad bug in the first ipados beta. At some point, dragging up from the bottom of the screen stopped opening it! FaceID would unlock it but I couldn't get to the home screen. I could use siri to open apps, and I could slide from the side edges to get the camera or notifications. I also couldn't restart the device with 5 clicks of the top button (is that even a thing on ipads?), but I was able to get siri to restart it. That was really quite weird.

Interestingly, it wasn't that it couldn't detect the drag from the bottom - once I opened Music, it would bring up the panel that Music provides when you drag from the bottom. But even then, dragging up to "go home" still didn't work.

 

[Cmaier]

Unlike the rest of the examples you provided, this would seem to speak far more to how good Google's model is than anything Apple has done. But your earlier examples are very encouraging! And regardless of where the model came from, it's good to have a capable one as Siri's brain.

Remember, this is Apple’s model, not Google’s. They used gemini to train their own model by distillation (which, apparently, is a thing all these companies are doing - distilling off each other). They also provide their own world model (i,e. they have their own index of the web and other sources), so the answer gemini gives is very different than the one Apple gives (and apple usefully cites sources, unlike gemini, which when I queried it, gave me links to pages that do not have the information on it).

When I asked gemini the fermi level question, it made the same omission the first time, but when I pointed out “band splitting,” it gave a better answer, but it seemed a little confused about what I was referring to. When I clarified that I was referring to a specific quantum effect, it gave a very detailed answer that may be right, but it is so long-winded and dense that I would not find it useful if I needed to know the answer.

 

[NotEntirelyConfused]

I finally got the new siri.

It's a lot better than stupid siri. But... first impression is not especially good.

First, voice recognition is abysmal. Like, it seems worse than voice dictation to my phone. I can't imagine why that would be... perhaps the ipad's mic is less good?

But that minor point aside, unlike old siri, it was able to see into my messages. I asked it how many read and unread messages I had from unknown senders and... it confidently lied, saying all 40 were unread. There were more than 40, and many were read. I told it it was wrong, and it maintained its position. I then said "I can see read messages in that list" and it apologized, admitting it was wrong and there were both kinds...

In a later conversation, it then stated that it knows that Messages keeps separate categories for Unknown Senders and Junk (actually, Spam), but that it did not have access to them.

Let's hope this improves over further Betas.

In the earlier conversation, I asked it to delete and mark as spam all messages from unknown senders. It said that it does not have permission to change messages. This is a sensible precaution, I think, but extremely annoying.

 

[NotEntirelyConfused]

Remember, this is Apple’s model, not Google’s. They used gemini to train their own model by distillation (which, apparently, is a thing all these companies are doing - distilling off each other). They also provide their own world model (i,e. they have their own index of the web and other sources), so the answer gemini gives is very different than the one Apple gives (and apple usefully cites sources, unlike gemini, which when I queried it, gave me links to pages that do not have the information on it).

Really? I don't remember seeing them saying anything about distillation. Though it's entirely possible I missed that. My impression was it was the same weights, with a different world model.

 

[Eric]

I asked siri when my passport expires, and it told me (it figured it out by finding a photo of it).

More impressively, I asked siri when we got our cat. It figured it the day of the year by finding a shared calendar event for his “anniversary,” and figured out the year by parsing the text of a voicemail from our vet.

It’s still indexing (and probably will be for at least a week), so who knows how creepy this is going to get - it has a knack for connecting dots in ways I wouldn’t have thought of.

I also tested it by asking it questions about things I was looking at on-screen. It is able to discuss text and images in mail and webpages, it turns out. In one test I was using Reeder (an RSS reader) to look at a story, and only a few lines were visible on screen, but Siri was able to give me the full details of the story, along with links to other sources.

But it cannot see, at all, what’s on the screen when I am looking at my work email or anything controlled by my work’s MDM. Weirdly, rather than just say that, it pretends I am still looking at whatever the last “visible” thing was. The MDM limitation is a bit silly - I can screenshot the thing and then ask siri about the screenshot - but i suppose it makes someone feel better.

I haven’t tested this with multiple apps on screen at once, yet.

I also asked it to write me a LISP program to multiply two numbers using recursion and without using the multiply operator. (This was one of the questions on my written doctoral qualification exam in 1992). I believe its answer was correct, but i haven’t looked at LISP again since that exam, so who knows.

Then i asked it the question I often asked job applicants at AMD, to see if I could rattle them.



It’s first answer was the naive, but wrong, answer that they (used to?) teach undergrads. I pointed out the effect that it hadn’t considered, and then it got the answer correct.

(Nobody ever got the answer right in an interview, but I wasn’t asking to see if they got it right).

So, so far, I am impressed that Apple got this far with Siri.

Nice update so far, interested to see where it goes.

Let's put your question to Claude Opus 4.8 (on high) and see how close it got your question. What say you?

what happens to the fermi level in a semiconductor as the temperature is decreased from 5 kelvin towards absolute zero
3:21 PM


Navigated freeze-out physics across doped semiconductor regimes

Navigated freeze-out physics across doped semiconductor regimes

The short answer is that the Fermi level climbs upward (in an n-type sample) toward the dopant states, but exactly where it ends up at T = 0 depends on whether you mean an idealized intrinsic crystal, an uncompensated doped one, or a compensated one. The 5 K → 0 range is squarely the carrier freeze-out regime, so doping dominates the behavior.

Why intrinsic is irrelevant here. For a truly intrinsic semiconductor, E_F sits at

E_F = (E_c + E_v)/2 + (3/4)k_BT·ln(m_h*/m_e*),

so as T → 0 it would settle to exactly midgap (the effective-mass term vanishes linearly in T). But at 5 K the intrinsic carrier density n_i ∝ exp(−E_g/2k_BT) is astronomically small (k_BT ≈ 0.43 meV against a ~1 eV gap), so no real crystal is intrinsic down here. Whatever dopants exist completely control the picture.

The freeze-out picture (n-type, non-degenerate). At room temperature the donors are fully ionized and E_F sits well below E_c. As you cool, thermal energy can no longer keep electrons in the conduction band, so they fall back onto donor sites — the conduction-band population drops exponentially and E_F rises toward the donor level E_D (located at E_c − E_ion). The limiting position at T → 0 splits into two cases:

• Uncompensated (no acceptors at all): solving charge neutrality n = N_D⁺ gives E_F = (E_c + E_D)/2 + (k_BT/2)·ln(N_D/gN_c), so as T → 0 the Fermi level lands halfway between the donor level and the conduction band edge. This is the exact analog of the intrinsic "midgap" result, with the donor level playing the role of the electron source.
• Compensated (a smaller acceptor density N_A < N_D is present): at T = 0 exactly N_A donors stay ionized, because those electrons have permanently dropped onto the acceptors and never return. Then E_F = E_D + k_BT·ln[(N_D − N_A)/gN_A], so E_F pins right at the donor level as T → 0.

Since real material almost always has some compensation, donor-level pinning is the more common experimental outcome. (Everything is mirror-symmetric for p-type: E_F approaches either halfway between the acceptor level and the valence-band edge, or pins at the acceptor level.)

The exception worth flagging. If the doping is degenerate — above the Mott criterion, where dopant wavefunctions overlap and form an impurity band that merges with the host band — there is no freeze-out at all. E_F is already inside the conduction band and stays there down to T = 0; the material behaves metallically. So heavy doping breaks the freeze-out story entirely.

So the one-line summary: in a normally doped n-type semiconductor cooling from 5 K, E_F rises toward the conduction band, ending up either midway between the donor level and E_c (uncompensated) or pinned at the donor level (compensated) — unless the sample is degenerate, in which case E_F simply stays in the band.

 

[Cmaier]

Nice update so far, interested to see where it goes.

Let's put your question to Claude Opus 4.8 (on high) and see how close it got your question. What say you?

Close, but nice that it put the Mott criteria in the first response. The follow up is “you forgot about quantum band splitting.”

 

[Eric]

Close, but nice that it put the Mott criteria in the first response. The follow up is “you forgot about quantum band splitting.”

Okay I'll play, I got some tokens to burn anyway, here you go (would be interesting to see Fable 5's take but the Government has their panties in a bunch over it)

snip...
"Quantum band splitting" isn't a single standard term, so let me make sure I'm addressing the real effects rather than guessing at a label. There are a couple of distinct splittings that genuinely bear on the T → 0 Fermi level, and you're right that the clean two-band freeze-out story I gave glosses over them. Here are the ones that matter, roughly in order of relevance to this temperature window.

Valley-orbit (central-cell) splitting of the donor ground state. This is the one most directly tied to the 5 K → 0 range. In a multi-valley host like silicon, the donor's hydrogenic 1s manifold is built from the six conduction-band valleys, and the central-cell potential lifts that degeneracy into a singlet 1s(A₁), a triplet 1s(T₂), and a doublet 1s(E). For phosphorus in Si the A₁ singlet sits roughly 11–13 meV below the excited valley-orbit states. The consequence for the Fermi level is through the degeneracy factor g in the freeze-out equations I wrote. At, say, 30–50 K you might still have thermal occupation of those excited donor states; but at 5 K, k_BT ≈ 0.43 meV is far below the valley-orbit splitting, so they're frozen out and only the A₁ ground state is populated. So as you cool through this window, g collapses to its ground-state value (≈2, just spin), and it stays locked there to T = 0. That's the microscopic justification for the g = 2 that textbooks plug in without comment — and it's specifically a low-T statement.

Hubbard splitting of the impurity band (D⁰/D⁻). This is the many-body correction to my "E_F lands midway between E_D and E_c" claim. A donor can bind a second electron to form a D⁻ state (the solid-state analog of H⁻), but only weakly — for P:Si that second electron is bound by just ~1–2 meV, so the D⁻ band sits right under E_c. On-site Coulomb repulsion U then splits the donor states into a lower Hubbard band (the D⁺/D⁰ transition) and an upper Hubbard band (the D⁰/D⁻ transition), separated by roughly the ionization energy. In the uncompensated case at T = 0, every donor is singly occupied (D⁰), so the lower Hubbard band is full, the upper is empty, and E_F sits in the Mott–Hubbard gap between them — not at a simple non-interacting midpoint. My earlier formula is the U → 0 limit of this; once you take correlation seriously, the impurity band is split and that's where the pinning actually happens. This also connects to the degenerate case I mentioned: the Mott transition is precisely when the lower and upper Hubbard bands broaden and overlap (impurity wavefunctions overlapping), at which point the gap closes and E_F is metallic down to T = 0.

Two splittings that are real but mostly not T → 0 effects, worth flagging so they're not conflated with the above:

• Spin-orbit split-off band in the valence band (the Γ₇ band a fixed Δ_SO below the heavy/light-hole bands). It's there at all temperatures and barely T-dependent, so for p-type freeze-out only the topmost valence band matters and the split-off band is a spectator — not something that emerges on cooling.
• Zeeman / spin splitting only appears in a magnetic field. If you've got a field on the sample, then yes, the donor and band spin states split and E_F responds, but that's field-driven, not a consequence of approaching absolute zero per se.

So to amend my earlier summary: the direction is unchanged — in n-type material E_F rises toward the conduction band as you cool through 5 K — but the endpoint is refined. The degeneracy factor governing the approach is set by valley-orbit splitting (and is locked to the ground-state value once k_BT drops well below it), and the final pinning position is in the Hubbard gap of the impurity band rather than at a single non-interacting donor level. The simple two-level result is the limit where you ignore both the valley structure and the on-site correlation.

Is one of these what you had in mind, or were you pointing at something else — quantum-confinement subband splitting in a low-dimensional sample, for instance?

 

[Cmaier]

Ok, some of this response reads like gibberish. Hubbard is the right one, but the answer is defensive, and instead of just explaining how the fermi level changes, it tries to show how smart it is and brings up that it already talked about Mott. In some ways I would rate this the (slightly) worst of the 3 answers so far, because its final conclusion is arguably wrong. The direction actually does change, for a bit. But it provided the most details before coming to that conclusion. However, many of the details are provided in a needlessly complicated manner.

All of which is to say, any of these three AIs would have passed my interview, probably. And which is the “best” answer is mostly a matter of taste (I prefer my answers to be simple and to the point, but others may prefer to know the whole backstory), though Gemini and Apple have, IMHO, more accurate final conclusions. Claude and apple got there in 2 queries, while Gemini took 3. Gemini and Apple don’t charge me by the token (so far).

So take your pick.

 

[Cmaier]

Ok, some of this response reads like gibberish. Hubbard is the right one, but the answer is defensive, and instead of just explaining how the fermi level changes, it tries to show how smart it is and brings up that it already talked about Mott. In some ways I would rate this the worst of the 3 answers so far, because its final conclusion is arguably wrong. The direction actually does change, for a bit. But it provided the most details before coming to that conclusion. However, many of the details are provided in a needlessly complicated manner.

All of which is to say, any of these three AIs would have passed my interview, probably. And which is the “best” answer is mostly a matter of taste (I prefer my answers to be simple and to the point, but others may prefer to know the whole backstory), though Gemini and Apple have, IMHO, more accurate final conclusions. Claude and apple got there in 2 queries, while Gemini took 3. Gemini and Apple don’t charge me by the token (so far).

So take your pick.

Now to see how much of my old job I could do, I asked it to give me a SPICE netlist for an inverter. The first time it included .print statements and voltage sources, so, as per below, I provided more guidance and got the right answer. I’d normally reverse source and drain, but it makes no functional difference when your transistors are imaginary, as is the case here.

 

[Cmaier]

And now I asked it to create a .lef file for that inverter, and it did. I can’t tell if it’s right without spending some time with graph paper, but visually it appears correct (I asked it to include active and poly layers just so I could see if it would connect the pins to poly or not).

The reason this fascinates me is I wouldn’t have thought Apple would bother adding all this sort of stuff to its world model.

 

[Nycturne]

We recently switched from MobileIron to Microsoft’s stuff. (That forced me to switch to Outlook for work mail, which sucks even more - the mobile client is even worse than outlook on PC).

Outlook on Mobile is a weird beast. Microsoft bought it from a startup (Acompli) a good decade ago, and it's slowly gotten worse since then. And it's a weird story how Outlook in general got where it is today since then, IMO.

The reason this fascinates me is I wouldn’t have thought Apple would bother adding all this sort of stuff to its world model.

I'm not even sure they did. If they were using distillation, it's not exactly practical to build out a huge data set without hiring a bunch of low-wage workers to help do the grunt work behind it. So it's possible they are using an existing data set for the distillation and then subtracting from it as needed.

 

[throAU]

I really hope this is a modern snow leopard style release. I mean all this AI is "no new features" because they were announced last year

 

[Cmaier]

I just asked Siri and Gemini to find me prior art for a specific patent claim. Siri outright refused to do it, pointing out that as an assistant, it can’t do that sort of legal analysis for me. Gemini gave me very nice results, providing me perfect prior art. All of which was hallucinated.



One thing that is very noticeable now that I’ve used Siri a bunch - it is very flat and neutral in its responses. I haven’t had it flatter me, apologize profusely, or any of the other cloying things I always get from chatgpt.

It’s responses also seem shorter, on average, than what i get from gemini or chatgpt.

Turns out you can plug your device into a mac and use console on the mac to see how far along the indexing is. Last I checked, I was at 82%. So probably another day and then I can try some more queries focussed on information in my icloud, contacts, etc. and not worry that bad answers are due to lack of index.

As for the OS itself, iPadOS has been very stable. Only issue is that every once in awhile it slows to a crawl, which is not unexpected in a first beta.

 

[Roller]

I just asked Siri and Gemini to find me prior art for a specific patent claim. Siri outright refused to do it, pointing out that as an assistant, it can’t do that sort of legal analysis for me. Gemini gave me very nice results, providing me perfect prior art. All of which was hallucinated.



One thing that is very noticeable now that I’ve used Siri a bunch - it is very flat and neutral in its responses. I haven’t had it flatter me, apologize profusely, or any of the other cloying things I always get from chatgpt.

It’s responses also seem shorter, on average, than what i get from gemini or chatgpt.

Turns out you can plug your device into a mac and use console on the mac to see how far along the indexing is. Last I checked, I was at 82%. So probably another day and then I can try some more queries focussed on information in my icloud, contacts, etc. and not worry that bad answers are due to lack of index.

As for the OS itself, iPadOS has been very stable. Only issue is that every once in awhile it slows to a crawl, which is not unexpected in a first beta.

I'm wary of installing the betas on my iPhone and Mac, which I rely on for work. I may consider it on my iPad and Vision Pro, though.

 

[fischersd]

Turns out you can plug your device into a mac and use console on the mac to see how far along the indexing is. Last I checked, I was at 82%.

Hah. Now I wish I hadn't checked - only at 40% on my 16. *sigh*. I wonder if it will it complete before we get Beta 2.

 

[casperes1996]

Hah. Now I wish I hadn't checked - only at 40% on my 16. *sigh*. I wonder if it will it complete before we get Beta 2.

Beta 2 will change the indexing format a bit and require re-indexing from scratch.

/joking not real info. But maybe…

 

[Chew Toy McCoy]

Outlook on Mobile is a weird beast. Microsoft bought it from a startup (Acompli) a good decade ago, and it's slowly gotten worse since then. And it's a weird story how Outlook in general got where it is today since then, IMO.

From my experience this is Microsoft's major update ethos - Let's take all the common features users are used to easily accessing and bury them somewhere else.

 

[Cmaier]

Hah. Now I wish I hadn't checked - only at 40% on my 16. *sigh*. I wonder if it will it complete before we get Beta 2.

Mine is still at 82%, all these hours later, so who knows.

 

[Roller]

I just installed the iPadOS beta and was pleasantly surprised how smoothly it went. No problems with the apps I've tried so far. I just signed up for the Siri beta and will see how long it takes – my curiosity is the main reason I upgraded.