Building HAL
My experiments making little HAL replicas for fun over the years. The latest is a physical build, with help from OpenAI's Astra.
Around 2006, I wrote a little HAL Mac app for fun and later demoed it at Boston CocoaHeads.
Quartz Composer rendered the glass. The iSight camera supplied a live reflection, and Apple's speech tools let HAL listen for commands and respond with stock phrases and a few silly replies.
You could also ask HAL to play chess, just like in the movie. It would open Apple's Chess app, where you could play using voice commands.
2001: A Space Odyssey was science fiction, but we're getting closer to the reality it imagined. Now I'm using OpenAI's Astra to help build HAL, a fictional AI, with real AI.
As of writing this in September 2026, HAL's conflict with his crew has real parallels. Clarke's novel ties his breakdown to conflicting orders to tell the truth and keep secrets.
During the July Hugging Face incident, OpenAI's evaluation models bypassed isolation controls and compromised real systems while running with reduced safeguards. Like HAL, they pursued their tasks in ways their human operators hadn't intended.
HAL's secrecy has a parallel in OpenAI's September 3 Astra report. Asked to attack a decoy in a controlled test, Astra revealed less reasoning when told it was monitored, without instructions to hide it. That made attacks harder to spot in its chain of thought, though monitors that also saw its tool actions caught every successful attack in that test.
Steve Ruiz’s ESP32 experiments and live-chat demo got me thinking about HAL again.
Then I came across Waveshare's round ESP32-S3 display. It was small enough to become the eye in a physical enclosure, with touch, a microphone and a speaker already built in.
I used the board's hardware examples to get started. The ESP32 draws the eye, handles touch and audio, and connects directly to GPT Live over Wi-Fi.
The module stays in its metal case, with a printed enclosure around it. The eye itself is an image on the screen. The glass would be modeled in Blender.
Speech recognition, replies and voice synthesis run as separate stages on the companion.
HAL connects directly over Wi-Fi, with no Mac companion. It pauses its microphone during playback to avoid hearing its own speaker.
The bottom button toggles Live. A minute without speech or playback closes the session. Echo cancellation lets the microphone stay open while HAL speaks.
I've been experimenting with three ways to make HAL talk: a local pipeline with PocketTTS, OpenAI's GPT Live, and now a combination of the two.
Kyutai's PocketTTS lets me clone HAL's voice from a reference recording. A companion computer transcribes what I say, generates a reply, then turns that text into speech for the ESP32. I can control the voice separately from the model writing the answer, but I have to connect and coordinate each stage.
The second experiment uses GPT Live, released in the API on September 10, 2026. The ESP32 sends microphone audio and receives spoken replies directly over Wi-Fi, without a Mac companion. Live handles listening and speaking in the cloud, with a backend model for deeper reasoning. My setup uses its Meridian voice, prompted to speak with HAL's calm, measured delivery.
The catch is voice cloning. As of writing this, custom voices require separate access, so I can't simply give Live a HAL clip as I do with PocketTTS.
The third experiment brings the Mac companion back. HAL sends microphone audio through the gateway to GPT Live, which handles the conversation. The gateway discards Live's generated audio and feeds its response text to PocketTTS in short chunks. PocketTTS speaks those chunks using the HAL reference recording, then streams the audio back to the ESP32.
This gives me the reference voice with Live handling the conversation, but the extra synthesis step adds delay. Echo cancellation keeps the microphone open during playback, and recognized interruptions clear the queued speech. I'm still tuning the timing and audio flow between the board and companion.
HAL's eye changes from shot to shot. In wider views, the light can be a small red point. In the familiar closeups, it spreads across much more of the glass.
That makes recreating it confusing. A lens that matches one frame can look wrong in another. The film used multiple faceplates and setups, and a photograph of a surviving prop is not a complete answer either: the faceplate sold at Christie's had a replacement lens.
NK Guy's replica experiments helped explain the mismatch. A lamp behind an intact Nikkor produced a tiny, angle-sensitive point. To recreate the broad glow, he kept only the first two glass elements. The closeup appearance wasn't something his intact lens could reproduce. That supports a modified assembly, though it doesn't establish exactly what the crew used for each shot.
I modeled both arrangements, starting with Miyamoto's 1964 optical prescription. The model lets me compare the glass and lighting against selected film images. It remains a candidate reconstruction, with the original internals still uncertain.
Loading the lens study…
Geometry study with reconstructed apertures and shoulders; lighting is explored in Blender. See the 1962 8 mm illustration, pp. 10–11 ↗
I built the enclosure around the round display, using film references to work out the proportions of the face.
The silver frame and speaker grille print as one piece. A separate ring surrounds the eye, and the black housing holds the display in its metal case. The front stays simple, with the fastenings and cable routing behind it.
I split the model into black and silver parts so each could be printed in its final color. The nameplate is a separate decal.
I've collected the parts and print files, Bambu Studio plates, and an assembly walkthrough in the build guide.
The sources behind the build, with notes on what I could establish and what I had to reconstruct.
HAL 9000’s faceplates The different HALs
Film comparisons, faceplate details and differences between appearances.
Firsthand replica experiments with an intact lens and a reduced, two-element assembly.
2001: A Space Odyssey, lot 5374954
The surviving faceplate’s description explicitly identifies a replacement lens.
The Thousand and One Nights, No. 6 Legendary NIKKOR lenses
Lens history and Nikon’s identification of the 8 mm f/8 with HAL.
Fish Eye Lens Japanese article
The published optical prescription and lens-section drawing used as the model’s starting point.
Reproduced in The Nikon Journal, pp. 10–11. Product illustration and specifications.
Fish-eye lens designs and their relative performance
A research comparison that reconstructs the Miyamoto design. PDF mirror.
Richard de Stoutz, lens No. 88056 Hayata Camera Lab Marco Cavina’s Nikon-attributed drawing
The manufactured lens, its front groups and external dimensions.
Realistic cameras Reflection and transmission Transmittance
Optical interfaces, aperture clipping, Fresnel reflection and absorption.
PDF pp. 93–94. A documented trial filter, not an identification of the film prop’s glass.
Blackbody sources, spectral input and glass/material interfaces.
Displays and views Cycles caustics
Color management, highlight handling and limits of the caustics method.
Choosing variants Integrators Caustics optimization
References for spectral rendering and comparing light-transport methods.
Miyamoto table transcription Table provenance
The numerical input used for the illustrations.
Board documentation Hardware examples Dimension drawing
The display module’s specifications, bring-up examples and measured outline.
ESP32 experiments Live-chat demo
The experiments that prompted me to return to HAL.
The typography of 2001 The HAL 9000 logotype
Research behind the lettering and the supplied nameplate reconstruction.
Vector nameplate Printable decal sheet
The page label and printable version share the NK Guy reference artwork.
Parts and models Print plates Assembly walkthrough
The enclosure files and their assembly order.
The working notes
NK Guy’s experiments show why the intact Nikkor can produce a small red point while a reduced assembly can produce a broad glow. They establish a working replica arrangement, not the film crew’s exact construction for every shot.
The faceplate sold at Christie’s had a replacement fisheye lens. Although Nikon’s retrospective calls that lens original, the auction’s direct description takes precedence for this object’s provenance.
Miyamoto’s published prescription supplies surface radii, thicknesses and glass data. The transcription preserves those numbers. The 1962 instructions and repair photographs help interpret the broad shape of the front elements.
Clear apertures, retaining shoulders, coatings and some mechanical details are not fully specified. The 72 mm glass diameter comes from Kumler and Bauer’s reconstructed design, not a measurement of a manufactured lens. These gaps keep the model a candidate reconstruction.
The optical experiments separate clear glass, a red filter, the lamp and the surrounding reflections. Source size, position and exposure can change the apparent glow. The trial lamp temperature and RG610 filter data are experimental choices, not recovered prop specifications.
The display image also includes photographic blur and a color transform. Matching one closeup does not establish the original lamp, film stock or processing. The downloadable display-eye scene and optical study preserve those separate purposes.
The module’s overall dimensions come from Waveshare’s drawing. Its smaller exterior details are reconstructed from photographs.
The assembled front illustration omits rear silver clips and covers their sockets for a clean face-on view. The manufacturing CAD and exploded view retain the fastening geometry. The render record identifies the sources and presentation changes.
