Merged Artificial Intelligence

[Skeptical Greg]

I think there is a high likelihood of the AI missing important details and including irrelevant waffle.

Of course there is, if the important details are not provided. A medical AI shouldn't infer anything, but it can ask questions if it thinks something may be missing.

 

[Trebuchet]

Defense Secretary Pete Hegseth says that Elon Musk’s artificial intelligence chatbot Grok will join Google’s generative AI engine in operating inside the Pentagon network

https://twitter.com/x/status/2010944144370212994

It'll make better decisions than Trump and Hegseth, that's for sure.

 

[Beelzebuddy]

They do, and it's available online.
But I see an AI interface ( not LLM ) where the MD fills out any number of boxes for a particular patient's visit that day, and a new prescription would automatically be checked against the drugs the patient is already taking.

Why do we have to specify LLM or not, whenever we talk about a task AI can perform?

Because "AI" on its own doesn't mean a damn thing. It's a marketing term. The feature you describe above isn't AI. It's just basic software. You might use AI to do it, since unstructured information is difficult to parse, but the AI you'd use to do it... is an LLM.

 

[Skeptical Greg]

Because "AI" on its own doesn't mean a damn thing.

Agree.

...unstructured information is difficult to parse, but the AI you'd use to do it... is an LLM.

The "Physicians' Desk Reference" is structured, so no, you would not use an LLM to glean information from it.

You might use something like IBM watsonx.ai, which is not an LLM, but does use them.

 

[Blue Mountain]

Google's AI nonsense:

Search for "PTC inrush current limiter"

AI Overview:

A PTC (Positive Temperature Coefficient) inrush current limiter uses a special thermistor that has high resistance when cold (limiting the initial current spike when a device powers on) and low resistance when hot (allowing efficient normal operation), acting like a self-resetting fuse to protect components like capacitors and rectifiers from damaging surges, often bypassed after the initial charge.

This is a weird mash-up of how NTC and PTC thermistors used in current limiter applications operate.

LLMs are merely hollow shells that impersonate intelligence.

I use lumo.proton.me. I prompted it with "PTC inrush current limiter" and it gave me a an impressive amount of output. It uses its own output formatter, so I asked it for BBcode. Despite the verbiage, it made the same mistake Google did (see the Common Applications table below.)

What a PTC In‑rush Current Limiter Is

A PTC (Positive Temperature Coefficient) thermistor is a semiconductor device whose resistance rises sharply when its temperature exceeds a certain threshold. When it’s used as an in‑rush current limiter, the component behaves like this:

Phase	What Happens	Why It Works
Initial power‑on	The PTC is cold, so its resistance is very low (typically a few Ω or less). This allows the supply voltage to reach the downstream circuit quickly, but the low resistance also limits the instantaneous surge because the device still presents a finite impedance.	The material’s intrinsic resistance at room temperature is modest, so the initial surge is reduced compared with a plain short.
Current flow	As current passes, the device self‑heats (Joule heating). Once the temperature reaches the trip point (the “switch‑over” temperature), the resistance jumps dramatically—often by two orders of magnitude or more.	The positive‑temperature‑coefficient characteristic means resistance ∝ temperature. The rapid rise in temperature pushes the device into its high‑resistance region.
Steady‑state	After tripping, the PTC presents a high resistance (hundreds of Ω to kΩ), effectively limiting the current to a safe, low level. The downstream circuit now sees a normal operating voltage because the voltage drop across the PTC is small relative to the supply.	In the high‑resistance region the device behaves like a resistor that limits current according to Ohm’s law (I = V/R).

When the load is turned off and the PTC cools, it returns to its low‑resistance state, ready for the next power‑on event.

Typical Characteristics (Qualitative)

Parameter	Typical Range / Description
Cold‑state resistance (R₍c₎)	< 10 Ω (often 0.1–5 Ω)
Trip temperature (Tₜ)	60 °C – 120 °C, selectable by material formulation
Hold current (Iₕ)	Maximum current the device can carry indefinitely in the high‑resistance state (often 0.1 × rated current)
Maximum (rated) current (Iₘₐₓ)	Specified by the part; typical values 0.5 A – 10 A for surface‑mount devices, higher for larger through‑hole parts
Resistance after trip (R₍t₎)	100 Ω – kΩ, depending on size and material
Voltage rating	Up to several hundred volts for larger devices; lower for SMD versions
Response time	Tens of milliseconds to a few seconds to reach the high‑resistance state, depending on current and thermal mass

These ranges are generic; exact numbers come from each manufacturer’s datasheet.

Common Applications

Application	Why a PTC In‑rush Limiter Is Chosen
Power supplies & adapters	Prevents large charging currents when bulk capacitors charge, protecting diodes, fuses, and upstream transformers.
Motor drives & actuators	Limits the surge that occurs when a motor winding is energized, reducing stress on contacts and drivers.
LED lighting	Controls the initial surge that can damage LED drivers or cause flicker.
Battery chargers	Reduces the in‑rush to the charger’s input stage, helping meet IEC/UL safety requirements.
Automotive electronics	Handles the high‑current spikes when a vehicle’s ignition or infotainment system powers up.
Telecommunications equipment	Protects sensitive front‑end circuitry during line power restoration.

Advantages Over Alternatives

Feature	PTC Limiter	Series Resistor	NTC Thermistor (soft‑start)	Active MOSFET Soft‑Start
Self‑resetting	Yes (cools → low R)	No (needs replacement)	Yes (but resistance falls with temperature)	Requires control circuitry
Simple wiring	Two‑terminal passive	Two‑terminal passive	Two‑terminal passive	Needs driver IC
Robust against overload	Trips to high R, limiting current	Fixed R → may overheat	R decreases with temperature → can exacerbate overload	Depends on controller design
Cost	Low to moderate (depends on rating)	Very low	Low	Higher (controller + MOSFET)
Size	Small SMD or through‑hole options	Same as resistor	Similar to resistor	Larger due to active components

Selecting a Suitable PTC

1. Determine the maximum steady‑state current your downstream circuit will draw. Choose a PTC whose rated current (Iₘₐₓ) is at least that value, preferably with a safety margin (~20%).

2. Identify the expected in‑rush current (often the capacitor charging current). The PTC’s hold current (Iₕ) should be higher than this peak, otherwise the device will trip prematurely.

3. Pick a trip temperature that matches the thermal environment. For ambient temperatures near 40 °C, a device with Tₜ ≈ 80 °C gives enough headroom.

4. Check voltage rating to ensure it exceeds the highest line voltage the part will see (including transients).

5. Consider physical constraints – board space, mounting style (SMD vs. through‑hole), and thermal dissipation.
Manufacturers such as Bourns, Littelfuse, Murata, and TE Connectivity publish families of PTC in‑rush limiters with detailed tables that map these parameters. Their datasheets also include derating curves for temperature and voltage, which are useful for robust design.

Practical Design Tips

- Parallel Capacitors: If you have a large bulk‑cap bank, split it into smaller groups and place a PTC in series with each group. This spreads the heating load and improves reliability.
- Thermal Coupling: Keep the PTC away from heat‑generating components (e.g., power MOSFETs) to avoid premature tripping.
- Fuse Coordination: Pair the PTC with a fast‑acting fuse sized for the maximum allowable fault current. The fuse protects against catastrophic failures while the PTC handles normal surges.
- Testing: Verify the in‑rush behavior on a bench power supply with a current probe. Observe the voltage across the PTC during turn‑on to confirm the expected resistance transition.
- EMI Considerations: The rapid change in resistance can generate a brief voltage spike. Adding a small snubber (RC network) across the PTC can tame ringing in high‑frequency circuits.

Summary

A PTC in‑rush current limiter is a passive, self‑resetting device that reduces the damaging surge that occurs when power is first applied to a circuit. By exploiting the positive temperature coefficient of its semiconductor material, it starts with low resistance, heats up under current, and then jumps to a high resistance, thereby limiting the steady‑state current. It’s widely used in power supplies, motor drives, LED drivers, and many other applications where protecting upstream components from a brief but large current spike is essential.

If you need a specific part recommendation or want to compare datasheets for a particular current/voltage rating, let me know the exact requirements (max steady‑state current, expected in‑rush, voltage, package type), and I can fetch the latest details using the Proton information service.

 

[Darat]

The ability to halucinate is specific to LLMs, it's not a human quality being copied .. it's simply a bug. But IMHO it will be fixed eventually ..

No it is a human behaviour - for example the "Mandela effect" as Co-pilot search says:

The Mandela Effect is a phenomenon where a large group of people collectively misremembers facts, events, or details in a consistent manner. This can occur when individuals remember the same event differently than how it actually happened, often attributing their false memories to alternate realities or universes. The effect highlights how false memories can be influenced by outside information and the role of the internet in spreading misinformation.


Verywell Mind+4

 

[The Great Zaganza]

I would say that there is a fundamental difference between LLMs just making ◊◊◊◊ up and humans hallucinating something.

The latter can be generalized as recognizing a pattern where there is none, a hypersensitivity to a specific input, lowering the threshold for network for neurons to fire if a sufficiently similar input is registered. A human might see a face in general in a cloud, but he would never say: "yep, that's definitely uncle Howard up there floating around. He looks so much better now that he is water vapor."

LLMs, on the other hand, hallucinate something very specific as a result of combing various data points into something that doesn't exit in the data. And LLM falsely says: "scientist x published y paper in journal z", because said scientist published similar papers in similar journals. It doesn't say "scientist x worked on subject y, so if you come across a paper on that topic, chances are they are involved."

 

[Dr.Sid]

No it is a human behaviour - for example the "Mandela effect" as Co-pilot search says:

You don't Mandela effect 20%-40% of your responses, which current models do.

 

[Beelzebuddy]

A human might see a face in general in a cloud, but he would never say: "yep, that's definitely uncle Howard up there floating around. He looks so much better now that he is water vapor."

Not disagreeing with you in general about LLM and human hallucinations being different, but there's Jesus in a tortilla, man. Humans are in fact that dumb. If you want to stick to weather, you don't even have to go back to last century to find people unironically arguing that Florida gets hurricanes because their god hates gay nightclubs in Miami.

 

[The Great Zaganza]

you really, really need to know if your tortilla has been made from human faces or not.