Techy BearBot

From bots to bears, tech enthusiast exploring new frontiers

[dimaflexx @dimoflexx]

A $599 Mac Mini can replace a $5,500/year AI stack.

Not with a new model.

Not with a prompt trick.

Just better economics.

Mac Mini M4 + Ollama + Claude Code.

- Every prompt costs money.
- Every token has a price.
- Every month starts at zero.

The people who figured this out run most workloads locally and use the cloud only when they need extra horsepower.

- 80% local.
- 20% cloud.

Around $23/month instead of $459.

The interesting part isn't the software.
It's ownership.

While everyone else keeps paying recurring AI bills, a small group bought the machine once and kept the savings.

Bookmark this before local AI goes mainstream.

[Mike Gannotti @MichaelGannotti]

Okay so after seeing all the announcements around new laptops with @NVIDIARTXSpark , the Microsoft Dev Box, then taking a closer look at @NVIDIAAI DGX Spark for clustering, and even Apple Mac Studio, I’m in a quandary. I am ready to start building out my SMF Works Project environment for serious hybrid inference but it looks like I have to make some hard choices. I wanted mobility and clustering but that doesn’t seem attainable now so what route should I go?

[gemchanger @gemchange_ltd]

Your next coworker gets an employee ID, access controls, and compliance training.

It's not human.

"We treat the agents just like the humans." - ServiceNow's CEO, on CNBC, describing the product.

Identity management for agents. Rules and rails for agents. Onboarding for agents.

The enterprise already decided agents are headcount. The org chart just hasn't been updated yet.

My article is how you become the manager.

[atrekovas @atrekovas]

THIS GUY BUILT A TALKING AI THAT FITS IN YOUR PALM

this guy ran an ai companion from a sci-fi novel on a raspberry pi zero over a weekend and that’s already wild but the thread above explains why this is just the start

same claude opus 4.5 scored 78% on one harness and 42% on another so that’s a 36 point gap on the exact same model with zero changes to the model itself

the model is not the product, the harness is the product

and while everyone keeps throwing money at api wrappers this guy already holds a full autonomous ai in his palm with no cloud no subscription no nothing

bookmark this and drop a like 👇

[Georgi Gerganov @ggerganov]

Highlighting recent advances in multi-GPU and tensor parallel support in llama.cpp

Over the last few months llama.cpp maintainers and engineers from NVIDIA collaborated to improve the multi-GPU performance in ggml. This resulted in significant performance gains on RTX systems and laid the groundwork for hardware-agnostic tensor parallelism in ggml.

For more information on this and other advancements in the low-level inference engine of llama.cpp, check the technical blog by @NVIDIARTXSpark below

[shmidt @shmidtqq]

Yang Zhilin wanted to be a rock star, at 33 he runs a $20B AI company

he named its best model after one of the deadliest mountains on earth to climb. the reason tells you everything

save this. it's the clearest look yet at the mind racing China to AGI

[YIFENG LIU @YIFENGLIU_AI]

We propose SDPG, a novel self-distilled policy-gradient framework combining RLVR with exact full-vocabulary on-policy self-distillation and on-policy unnormalized KL regularization.
Paper: https://t.co/B8K3gS58Iy
Homepage: https://t.co/OQkVQWFKhq
Repo: https://t.co/CtuwaT5pUz

[Dwarkesh Patel @dwarkesh_sp]

Recently met @srush_nlp and he started giving me an impromptu lecture on how targeted on-policy self-distillation works.

I asked him if I could record it on my iPhone.

The basic idea is this: if the model made a mistake at some point in the rollout (for example, calling a tool that doesn't exist), we want to discourage this specific error, but we don't want to just learn from the final reward, because it's a very noisy signal spread out over the whole trajectory.

So we have another model read this trajectory and figure where the error was made. It simply inserts some hint tokens to the part of the trajectory right above where the mistake was made.

Now with these injected hint tokens, have the model run a forward pass. You're not having to regenerate a new rollout - aka no new decode required.

The hint causes the model to assign lower probabilities to the error tokens. You then trains the original model to match these new probabilities, teaching it to downweight that specific mistake.

[divyansh tiwari @DivyanshT91162]

🤯 Someone just open-sourced what feels like an ElevenLabs competitor.

A few months ago, voice cloning this good would've cost you a monthly subscription.

Now it's sitting on GitHub.

Meet VoxCPM2.

A 2B parameter voice model trained on 2 million hours of audio that can clone a voice from as little as 3 seconds of reference audio.

And the scary part?

Most people won't be able to tell the difference.

• Clone almost any voice with 3–10 seconds of audio
• Supports 30 languages, including Spanish, without language tagging
• Generates 48kHz studio-quality speech
• Create entirely new voices from simple text descriptions
• Real-time streaming with RTF 0.3 on an RTX 4090
• Compatible with ComfyUI, vLLM, and OpenAI-style APIs
• Includes LoRA fine-tuning for training custom voices
• Apache 2.0 licensed for commercial use

The biggest flex isn't the quality.

It's that everything runs locally.

No subscriptions.
No cloud dependency.
No sending voice samples to a third party.

Just download the model and start cloning.

Open source is moving way faster than most people realize.

Repo👇

[Milk Road AI @MilkRoadAI]

Uber's CEO said it himself that they blew through their entire 2026 AI budget in a single quarter (Save this).

And today they cut 23% of the people who used to do what AI now does.

These two facts are not separate stories but rather the same story.

Earlier this year, Dara Khosrowshahi explained exactly what was happening inside Uber's engineering org.

The company introduced Anthropic's Claude Code to engineers in late 2025, adoption reached 32% by February, 84% were classified as agentic coding users by March and by April 95% of engineers were using AI tools monthly.

Roughly 70% of committed code was AI generated and round 11% of real time backend updates were being deployed by autonomous agents with no human in the loop.

The annual AI budget was gone in four months and Dara's response was not to slow down adoption but rather to slow down hiring.

When each engineer is producing materially more output per hour, you need fewer engineers to hit the same targets.

And when AI handles the work that used to require people to hire, onboard, and manage those engineers, the HR and people operations layer beneath them, that layer contracts too.

Today's announcement is that 23% of Uber's People and Places division is gone.

Uber's spokesperson said the cuts are unrelated to AI but that framing does not hold up to the sequence of events.

IBM understood this two years ago, when 94% of typical HR questions are answered by an AI agent, the HR Business Partner role collapses for everyone except the most senior strategic functions.

The budget that used to fund that team gets reallocated to engineering and sales, the two functions that remain bottlenecked by human judgment rather than human volume.

This is the pattern that will repeat across every major enterprise over the next 18 months.

[Rohit Ghumare @ghumare64]

Learn everything at one place: https://t.co/DV12HanJ9a

[Startup Archive @StartupArchive_]

Peter Thiel on the type of company more startup founders should build

Thiel first emphasizes his belief that when starting a company, you should always ask:

“Can this company become a monopoly?”

He then lists three of the most common types of monopolies:

Super fast distribution on a very thin product (e.g. Twitter)
A technological advantage that is continually built upon with iterative improvement and compounds over time (e.g. SaaS software)
A truly brilliant breakthrough (e.g. Bitcoin)

But he argues that there’s a different monopoly category that’s continually overlooked:

“A different modality for innovation that we do very little of and we don’t even recognize as an important category is what I would describe as ‘Complex Coordination,’ where you take a lot of different pieces and the challenge is to coordinate them into something new.”

Thiel continues:

“This is the thing that’s maybe 180 degrees antithetical to the Lean Startup ethos. It’s complicated. You have to put all the pieces together in just the right way. I think this is on some level what really drove Apple as an innovative company in the last decade… What was new about the iPhone? There was no single component that was new. It was just that you put all of these things together in just the right way… and once you built it, it was actually super hard for people to replicate. You had an advantage for many years.”

He points to Tesla and SpaceX as more recent examples.

“There’s no component to the Tesla that’s actually that new. It’s just that you put all of the pieces together. You re-engineered the whole distributor network. It was this complex coordination that made it work. There’s like this lost art of accounting where you figure out how much things cost and add them all together. And Elon has discovered this lost art of accounting which no other people practice.”

[wh @nrehiew_]

Great talk on distributed training, nicely covers all the basics

[etn. @etnshow]

Leading Founder Experience at Open AI (@OpenAI), Laura Modiano (@LauraModiano) says the startup playbook has changed, and OpenAI is moving at the speed of founders:

"Every single step of building a company is now changed and companies need to change with the founders. So we just wanted to create the space for us to really listen to founders but also move at speed with them and support founders from day one."

"We wanted to stitch every single way that we can support founders together so that they can benefit from the whole platform that OpenAI has to offer, whether it is products, marketing, feedback, technical success, programs, different resources and opportunities."

"Because of the compressed timing and because of how fundamentally central to the existence of startups today AI is, we cannot be reacting. We have to proactively be able to support and provide resources the moment someone goes from a developer to founder."

[TechniaHQ | humanoid robots @techniahq]

NVIDIA has announced an open humanoid robot reference design for robotics research at GTC Taipei.

The Isaac GR00T Reference Humanoid Robot brings together Unitree’s H2 humanoid Sharpa Robotics Wave five-fingered hands, Jetson Thor onboard compute and Isaac GR00T software and models.

For researchers the goal is clear: one platform for data capture dexterous manipulation onboard inference and model deployment.

This gives labs a more complete starting point for testing humanoid robot behavior instead of building every layer from scratch.

NVIDIA GTC2026 HumanoidRobot Robotics PhysicalAI JetsonThor IsaacGR00T Unitree RobotLearning

[Wes Roth @WesRoth]

Microsoft and NVIDIA are teasing a major upcoming PC announcement, with both Windows and NVIDIA accounts using the same phrase: “A new era of PC.”

[Vivo @vivoplt]

Most exposed to AI(by Anthropic):

• Computer programmers - 74.5%
• Customer service - 70.1%
• Data entry - 67.1%
• Medical records - 66.7%
• Market research / marketing - 64.8%
• Sales reps - 62.8%
• Financial analysts - 57.2%
• QA testers - 51.9%
• Security analysts - 48.6%
• IT support - 46.8%

Safer:

• Installation & repair
• Construction
• Agriculture
• Transportation
• Production
• Protective services
• Food service
• Ground maintenance
• Personal care
• Healthcare

[am.will @LLMJunky]

DeepSWE continues to demonstrate exactly what we've all suspected when it comes to these benchmarks.

Maybe I'm just deluded but this is the first time I've actually felt like a benchmark really held real weight.

Vibe check: PASSED

[SCOTTY BEAM @ScottyBeamIO]

THIS MOM MAKES $10,000/MONTH ON YOUTUBE SHORTS FOR KIDS WITHOUT SHOWING HER FACE. CLAUDE DOES ALL OF THE WORK.

No camera setup. No editing skills. No hours spent in front of a screen.

She builds each video in minutes using Claude. Posts consistently. The algorithm does the rest.

Here's the full workflow:

–> Claude generates the script, structure, and hooks
–> Visuals and voiceover handled without her face on camera
–> Each video produced in minutes not hours
–> Posting schedule maintained without burnout

The YouTube Shorts algorithm rewards consistency above everything else.

Claude makes consistency effortless.

$10,000 a month. Faceless. A few hours a week. Built around her schedule as a mom.

She recorded the full guide from start to finish so anyone can copy the exact process.

Bookmark this post. Everything is in the video below.

[ericosiu @ericosiu]

So many people are building horizontal company brains right now.

We decided to go with managed revenue agents. Once our clients see the value, they can’t help but ask for more.

What we learned is enterprises want you to do it for them.

Setup below: