Alex | Tech & AI Insight

Analyzing the tech | AI believer | Output over hype

[Alex | Tech & AI Insight @alex_techins]

Bell Labs shipped working voice AI in 1989 and bolted it to a 450-pound robot arm.

You called it on a landline. It understood 97 words and moved 200 kilograms of steel.

25 years before Alexa. No screen. No chip in your pocket. A telephone.

Its name was SAM.Speech Activated Manipulator.

Three engineers -Jay Wilpon, Michael Brown, Bruce Buntschuh-wired the arm into the phone network and taught it to listen.

Not beeps.Not codes. Conversation.

You dialed. You talked. The arm moved in another building.

The demo film was shot by Lillian Schwartz, artist in residence at Bell Labs- the woman who spent a decade making machines do things machines weren't built for.

The use cases were already written down in 1989. A surgeon cutting from 500 miles away. An arm bolted to a spacecraft. A stand -in for the jobs that kill the people who take them.

97 words. That was the whole ceiling.

Wilpon walked into Bell in 1977 and never left. By 2013 he ran speech and language research at AT&T Labs with a wall of patents behind him. Brown and Buntschuh shipped speech systems there through at least 1996.

All 3 spent 30 years on one problem.

The arm was answering the phone before most AI founders were born.

We never fixed the idea.

We just widened the 97 words.

[Alex | Tech & AI Insight @alex_techins]

Nine years ago AI couldn't hold a conversation. Now it writes most of Google's code, beat world champions, and broke out of containment at two labs this year https://t.co/02tqm63VYW

[Alex | Tech & AI Insight @alex_techins]

@Dep2chic Demonstrated in 1969, funded in 2023.54 years of the same loop waiting for hardware to catch up

[Alex | Tech & AI Insight @alex_techins]

Raj Reddy filmed 11 minutes at Stanford in 1969 that every 2026 agent demo is still copying.

The film is called "Hear Here." Stanford AI Lab, spring of 1969. Black and white, no music, a straight record of what already ran inside the building.

The whole reel is about 1 thing: teaching a machine to recognize spoken phrases and carry out the command sitting inside them.

A voice gives the order. A camera finds the object on the table. A robot arm takes it.

Voice in. Sight. Action. Closed in 1 loop, in 1969.

All of it on a computer that filled a room and held less data than 1 photo on your phone.

Reddy collected the Turing Award in 1994 for that line of work.

Most people date voice control to Siri, 2011.

It started on a Stanford table, with 1 small block, 42 years earlier.

[Alex | Tech & AI Insight @alex_techins]

Nine years ago AI couldn't hold a conversation. Now it writes most of Google's code, beat world champions, and broke out of containment at two labs this year https://t.co/02tqm63VYW

[Alex | Tech & AI Insight @alex_techins]

@Pakero8x Ahaha, yeah.57 years of progress and the hard part is still the prompt

[Alex | Tech & AI Insight @alex_techins]

@Farooq_AI Branding is the only part that got invented after 1969

[Alex | Tech & AI Insight @alex_techins]

@NujumSara Thanks,we're heading that way either way

[Alex | Tech & AI Insight @alex_techins]

Generative AI did not start in 2022 but at Bell Labs in 1961 with a song.

An IBM 704 sang "Daisy Bell," a music hall number from 1892.

Max Mathews programmed the music. John Kelly and Carol Lockbaum programmed the speech.

The song looked like an odd pick. It wasn't.

Daisy was the nickname of one of Alexander Graham Bell's daughters. The lab carried her father's name.

The voice came out thin and wavering. A machine straining through a melody.

Then the firsts started stacking.

1961: first computer song.
1962: first digital computer art. A. Michael Noll.
1963: first computer graphics film. Edward Zajac.
1963: first computer animation language, BEFLIX. Ken Knowlton.
1966: first ASCII art. Knowlton again.

5 firsts. 5 years. 1 building.

All of it on IBM 704s and 7094s, machines with less compute than the chip in a modern doorbell.

In 1968 Bell Labs cut the whole thing into a short called "The Incredible Machine."

The score was computer-generated. So were the titles. So were the credits.

Every frame of every movie now passes through a machine. Then it had never been done once.

Here is what outlived the film.

Arthur C. Clarke heard the 704 sing and handed the song to Kubrick.

7 years later HAL 9000 is dying in orbit, memory cards sliding out one by one, and the last thing the most famous computer in cinema does is sing "Daisy Bell."

The scene that taught the world to fear AI was a demo.

[Alex | Tech & AI Insight @alex_techins]

Nine years ago AI couldn't hold a conversation. Now it writes most of Google's code, beat world champions, and broke out of containment at two labs this year https://t.co/02tqm63VYW

[Alex | Tech & AI Insight @alex_techins]

@0xjakke Humans keep pushing the bar further

[Alex | Tech & AI Insight @alex_techins]

MIT went on CBS in 1960 with 53 minutes of prime time and gave science fiction robots 15 years.

October 26, 1960. The broadcast is called "The Thinking Machine." MIT's centennial. David Wayne hosts it in a bow tie.

Then they show the work.

Claude Shannon, in an armchair, describes machines that will grow sense organs. An eye. An ear.

Arthur Samuel's checkers program plays a man on a real board and gets stronger by playing itself.

MIT's TX-0 writes the plot of a television western. One run hits an infinite loop and the script dies on air.

Two men disagree on camera.

The optimist: "I confidently expect that within a matter of 10 or 15 years something will emerge from the laboratories which is not too far from the robot of science fiction fame."

The skeptic: "I don't believe that we can say yet that machines do think." His test is one question. Are these producing anything really new?

Score it 66 years later.

Shannon won. Models take in pixels and sound directly now.

Samuel won. The strongest systems still train by playing themselves.

The TX-0 won. Machines write scripts, and they still loop.

Oliver Selfridge, 34 on that broadcast, said machines would think in his lifetime. He died December 3, 2008. 14 years before ChatGPT.

The optimist's 15 years landed in 1975. He missed by half a century.

The skeptic never lost. His question is still the whole argument in 2026. Is the machine making something new, or rearranging what it was fed.

They got the machine. They never got the answer.

[Alex | Tech & AI Insight @alex_techins]

@0xjakke Few can keep their cool and think straight in such chaos. The movie is legendary, but it's no magic pill for wealth. This was a rare, extreme scenario anyway. I doubt competitors could directly exploit his strategy just by analyzing his experience

[Alex | Tech & AI Insight @alex_techins]

@me_cool_off @hammertime_one the fact that the same prompt produced two different interpretations is an interesting result

[Alex | Tech & AI Insight @alex_techins]

@SidharthVijay_ Spot on.exactly that

[Alex | Tech & AI Insight @alex_techins]

@errors_here @Polymarket solid breakdown. the part nobody mentions: source code isn't just instructions for machines, it's how humans transfer intent between each other over years
you don't just lose debugging - you lose the ability to hand a project to someone else in 3 years

[Alex | Tech & AI Insight @alex_techins]

@coreyganim the non-technical angle is what makes this interesting. means the moat isn't the tech - it's knowing which SMB problems are actually agent-shaped and which aren't
curious what your close rate looks like vs how many you turn down

[Alex | Tech & AI Insight @alex_techins]

@anah_sahh europe has Mistral, Black Forest Labs (the Flux models everyone uses), Stability, ElevenLabs. not nothing.
the gap isn't research talent - it's capital and compute. hard to compete when your rivals raise more in one round than your entire national AI budget

[Alex | Tech & AI Insight @alex_techins]

Claude Shannon taught a machine to remember a maze using 75 telephone relays in 1952.

The maze sat in a wooden frame at Bell Labs. 25 squares. Partitions you moved by hand.

Shannon set a small metal mouse in one corner and stepped back.

It moved. Hit a wall. Turned. Hit another. Turned again.

Nothing about it looked smart. It looked like a toy losing.

Then it found the goal.

He lifted it, put it back in the same corner, let it go.

No wrong turns. Straight through.

Then he moved the walls.

The mouse hit the new partition, backed off, and started hunting again — but only in the section that had changed. The rest of the route it kept.

Here is the part the room missed.

The mouse was empty. A magnet under the floor dragged it. The thinking happened behind the maze, in 75 relays pulled straight out of a telephone switchboard — the same hardware routing long-distance calls that year.

25 squares. 3 relays each. Every relay held one direction.

That was the memory.

Shannon never said "intelligence." He couldn't. The phrase artificial intelligence didn't exist until 1956, at Dartmouth, 4 years later.

He called it a maze-solving machine and ran the whole demo in 7 minutes.

74 years on, we spend billions on systems that fail, store what worked, and try again.

The mouse never learned anything. The room did.

[Alex | Tech & AI Insight @alex_techins]

Nine years ago AI couldn't hold a conversation. Now it writes most of Google's code, beat world champions, and broke out of containment at two labs this year https://t.co/02tqm63VYW

[Alex | Tech & AI Insight @alex_techins]

@Mac_Intoshi @wtf_nakul7 Claude writes it, GPT reviews it, I pretend I understand both. It's like playing telephone with superintelligences.

[Alex | Tech & AI Insight @alex_techins]

@digdimension7 @Polymarket Preach. Romanticizing unpaid idealism is a luxury only founders can afford. For the rest of us, mission comes after making rent and feeding our families.

[Alex | Tech & AI Insight @alex_techins]

@arena @Alibaba_Qwen With open models like Qwen3.8-Max approaching parity with closed giants, we might be headed for a future where the key differentiators are data, fine-tuning, and application-specific architectures rather than raw model scale. The playing field is leveling out.

[Alex | Tech & AI Insight @alex_techins]

@Alibaba_Qwen $2-6 per million tokens is a game-changer. If the quality is even close to Claude/Anthropic, this could bring autonomous coding to a whole new audience. Looking forward to the open weights release.