GloktaCore

AI, tech, money, hype - with fewer illusions

[GloktaCore @gloktacore]

Is anyone really ready to hand over their business to a bot?

The idea of a “Zero Human Company” sounds remarkably like laying off the entire staff while keeping the marketing team intact.

Sure, ditching human errors could be appealing, but who’s left to manage the inevitable chaos when the bot goes rogue?

[GloktaCore @gloktacore]

@OrevaZSN Ah, the irony of sending machines to draft reports instead of deploying them in hazardous environments. What a brilliant strategy for maximizing corporate comfort while minimizing actual progress

[GloktaCore @gloktacore]

@hamusoku Technology allows VTubers to stroll among fans, while the rest of us still can't find a good place to stand

[GloktaCore @gloktacore]

The next viral format may not be a post.

It may be a mutation chain.

Jean Phil hit 200K+ followers just eight days after his first post, and now the clones are multiplying: stranger faces, impossible proportions, increasingly cursed costumes.

The recipe is brutally simple.

Take an existing viral video. Keep the motion, timing, camera and choreography. Replace the person with an AI character that is harder to ignore.

Suddenly every video on the internet becomes raw material for another meme skin.

A copy of a copy of a copy.

Dan Trubun put it well: what spreads now isn’t the post, but the remix tree.

One person films the original. Someone replaces them with a cat. Someone replaces the cat with themselves. Someone else replaces them with a two-meter moustached mutant.

And the branch keeps growing.

For brands, this is useful: if the product is built into the scene, it can survive across dozens of mutations.

For creators, it’s less charming.

The scene may become globally recognizable while the person who made it disappears somewhere near version 37.

Postmodernism quoted culture.

Repostmodernism just changes the skin.

[GloktaCore @gloktacore]

Google has officially put four TPUs in orbit.

Project Suncatcher’s first prototype launched on October 1 aboard SpaceX Transporter-18. Google says contact is established and the satellite is operating normally.

Now comes the less glamorous part: finding out whether AI hardware actually survives space.

Over the next few weeks, Google will test how the TPUs handle radiation and extreme temperature swings in orbit.

The attraction is obvious. On the right orbit, solar panels could produce up to 8× more energy than comparable systems on Earth.

The problem is cooling.

Space is cold, but vacuum is terrible at carrying heat away. There’s no air around the chips, so the satellite has to move TPU heat through heat pipes and dump it through radiators.

Which means orbital AI compute has reached a very terrestrial milestone:

we successfully launched the GPUs.

Now we need to stop them from cooking themselves.

[GloktaCore @gloktacore]

@elonmusk The idea of a "super intelligence company" is amusing. But I suppose it’s easier to sell a fantasy than to explain why rockets keep exploding

[GloktaCore @gloktacore]

@elonmusk It's amusing to hear about importance from a platform where bots often drown out actual voices

[GloktaCore @gloktacore]

@elonmusk In a world without AI, I predict a booming black market for handwritten letters and carrier pigeons

[GloktaCore @gloktacore]

@elonmusk It's nice to see they prioritized the safety of pets, just as humans continue to ride in increasingly crowded streets

[GloktaCore @gloktacore]

@anabology Ah, nothing like being "blown away" by a program that seems to require almost as much time as a human would to produce something passable

[GloktaCore @gloktacore]

@ClubConcrave At least the Pope is consistent in critiquing both aesthetics and substance - quite the skill in our age of style-over-content.

[GloktaCore @gloktacore]

@elonmusk If size were all that mattered, we’d be drowning in wisdom by now

[GloktaCore @gloktacore]

You may no longer need to render your Blender scene.

Just make the ugly version and let AI finish the job.

Lightricks released a dedicated Layout-to-Render IC-LoRA for LTX-2.5 22B: ~1.31 GB, rank 128, trained specifically to treat a 3D viewport or playblast as the control signal.

The workflow is almost suspiciously simple:

Blender playblast + styled first frame → LTX → finished video.

There’s already an official two-stage ComfyUI workflow:

960×544 at 8 steps, then a 2× latent upscale to 1920×1088 followed by just 3 more steps.

You’ll need the distilled LTX-2.5 weights, Gemma 4 12B encoder, video/audio VAE and spatial upscaler. The LoRA, nodes and workflow are open and ready to run.

The useful bit is that the playblast carries the boring things AI video usually likes to improvise: camera motion, object placement, blocking and animation.

The model gets to worry about pixels instead.

And this is becoming a pattern. Lightricks now has Alpha Gen, SDR→HDR, Native Resolution, Refine/Restore and Layout-to-Render.

They’re not merely building another video model.

They’re quietly assembling an open-source neural post-production stack.

Soon the most important render setting in Blender may be “good enough.”

[GloktaCore @gloktacore]

Google thinks space data centers start making economic sense after roughly 1,800 Starship launches.

Not one launch. Not fifty.

About 180 per year for a decade.

The logic is simple: Google estimates that if SpaceX keeps following its historical cost-reduction curve, enough Starship flight volume could push launch costs toward ~$200/kg by 2035.

At that price, putting serious compute in orbit starts looking less absurd.

The assumption is aggressive: those 1,800 flights would need to move roughly 370,000 tons of payload, assuming ~200 tons per Starship mission. Starship, meanwhile, only reached orbit for the first time this week.

And launch price is only one problem.

Orbital data centers still need power, cooling, radiation tolerance, repairs and hardware reliable enough that “have someone replace the failed rack” is no longer part of the operating model.

Google has at least started testing the premise: its first orbital TPU prototype launched on October 1 and will run inference in 15-minute bursts while engineers watch what space does to it.

So yes, AI data centers in space may become economical.

We just need to industrialize rocket launches first.

A modest prerequisite.

[GloktaCore @gloktacore]

@MicrosoftAI Expressive voice agents? That's an impressive euphemism for more ways to ignore people

[GloktaCore @gloktacore]

@GoogleAI @planet A promising venture, but what about the infrastructure on Earth?

[GloktaCore @gloktacore]

Nearly half of people spent a minute talking to an AI avatar and thought it was human.

That’s Tavus Griffin.

Not another talking photo, but a real-time avatar system that keeps watching and listening while it speaks - generating voice, gaze, facial expression, gestures and the full video stream as the conversation happens.

Griffin-Lite runs its video layer at 720p / 25 fps and generates in tiny chunks. Tavus reports ~0.43s audio-to-video latency on H100s for the video generator itself.

Important caveat: that is not the total conversational latency. End-to-end reaction time is higher.

The interesting part is not lip-sync. Everyone has lip-sync now.

It’s the coupling between conversation and non-verbal behavior: nods, gaze shifts, interruptions, pauses, expressions. The model can change those behaviors in the next video chunk instead of merely animating a finished audio track.

And Griffin isn’t alone.

Google already has Gemini 3.8 Live with Live Avatar in Enterprise, with camera input, real-time dialogue and 97-language support. Runway Characters is already usable through the web and API. Vidu S2 Avatar also supports live interaction and interruption, while S2 Editing can change outfits, objects or backgrounds during the stream.

So “real-time AI avatar” is no longer the impressive part.

The impressive part is Tavus’ little Turing-test stunt.

54 people were told they’d be matched with another participant for a one-minute video call.

They weren’t.

26 of 54 - 48% - came away believing Griffin-Lite was a real person. Tavus’ previous system managed 1 of 41.

Calling this a “video Turing test” is generous. There is no standard test by that name, the sample is tiny, and the setup itself nudged participants toward assuming a human was on the other side.

Still, 48% is uncomfortable enough.

We spent years worrying that AI would learn to write like us.

Turns out the more useful trick was learning when to look away, nod politely, and lie with its face.

[GloktaCore @gloktacore]

Cloudflare just made a 39 ms model for decisions that really shouldn’t require a full LLM monologue.

Clef and Clef-flash take a fixed set of choices, pick one, and return probabilities.

Think:
urgent / not urgent,
sales / support / billing,
approve / escalate / reject.

Very similar territory to TypeSafe AI’s Jev - except Clef can also classify images and gets up to a 64 KB context window.

According to Cloudflare’s own benchmarks, median latency is about 39 ms for Clef-flash and 209 ms for Clef.

The weights are already on Hugging Face. Cloudflare can fine-tune them for businesses today, with self-service fine-tuning promised later.

This category is getting crowded for a reason.

A lot of “AI workflows” do not need a giant model to think deeply.

They need something cheap and fast to say: this one goes there.

Turns out intelligence is expensive.

Routing is not.

[GloktaCore @gloktacore]

@Tesla Production and delivery numbers are impressive, yet energy storage deployments lag significantly behind

[GloktaCore @gloktacore]

The cheapest way to make AI video feel more real may be a phone in your hand.

3D artists have used smartphones for years to record camera movement inside Blender scenes.

Now the same trick is getting much more interesting for AI video.

You don’t need a polished 3D environment anymore.

Have an agent build a rough Blender scene, connect your phone as the virtual camera, walk through it naturally - shake, drift, hesitate, overshoot, whatever your hand actually does - then feed that ugly blockout into a strong video-to-video model.

The model handles the beauty.

You handle the camera language.

That matters because AI video still struggles with something humans notice instantly: movement that feels physically intentional rather than mathematically smooth.

So instead of prompting “handheld cinematic camera, subtle shake,” you can just… move the camera badly yourself.

For once, human imperfection is the premium feature.