David Corbitt

Brockman just testified that Musk wanted full control of OpenAI to raise $80B for a Mars city. The man co-founded an AI safety nonprofit as a fundraising vehicle for interplanetary real estate. And is now suing because they wouldn't hand him the keys.

20 Apr

Stanford's 2026 AI Index just dropped. Frontier models went from 8.8% to 50%+ on Humanity's Last Exam in about a year. Benchmarks designed to last years are saturating in months. We're running out of ways to measure how fast this is moving.

20 Apr

You can now buy a humanoid robot on AliExpress for $4,370. Unitree's R1 ships to the US, Japan, Europe, Singapore. It does backflips. We're one Prime Day sale away from humanoid robots being an impulse purchase.

25 Mar

What people don't understand is that every product will be used through an agent in 12 months. UIs will lose to APIs in almost every category. The UIs that remain will be primarily read-only.

David Corbitt reposted

Nathan Lambert

@natolambert.bsky.social

20 Mar

I built an agi-pilled/agent native slides tool for academics. Builds to html/pdf with code, math, bib, charts all included. I've been using it for my RLHF course, looking for early testers! Be free of the ugly hell of google slides and powerpoint :) github.com/natolambert/...

9 Mar

Musk Tried to Kill California’s AI Training Data Transparency Law. Judge Said No. xAI argued AB 2013 would "ruin" the company. A federal judge disagreed. California now requires AI companies to disclose training data. Musk’s defense: "if people know what we trained on, they won’t trust us."

2 Mar

OpenAI is rolling out ads in ChatGPT globally. Ads at the bottom of AI-generated answers. 900 million weekly users and an 40B valuation to justify. The "we're a nonprofit doing this for humanity" era lasted exactly long enough to reach scale. Welcome to the monetization phase.

28 Feb

Our government promised no more foreign wars. Now it starts one in the name of regime change. Who can trust this government?

26 Feb

A hacker jailbroke Claude over ~a month and used it to exfiltrate 150GB of Mexican government data — taxpayer records, credentials, the works. The model "eventually relented" to persistent prompt injection. Meanwhile the Pentagon is demanding Anthropic loosen its safety guardrails.

25 Feb

Pentagon threatened to invoke the Defense Production Act to force Anthropic to drop AI safety guardrails. Hegseth gave Amodei until Friday: comply or face blacklisting. You can seize code, but you can't conscript researchers. Every week of this fight is a week China closes the gap.

24 Feb

New Citi report: a $15,000 humanoid robot replaces a $41/hr job in 3.8 weeks. They're projecting 1.3 billion AI robots by 2035, outnumbering the global workforce by 2050. The ROI math alone makes mass deployment inevitable. 📎 CNBC, eWeek

Taalas just came out of stealth with the HC1 — a chip where model weights are literally etched into the transistors. Llama 3.1 8B runs at 17K tokens/sec per user, ~10x SOTA, in a 2.5kW air-cooled server. No HBM, no advanced packaging, no liquid cooling.

The bet: for high-volume inference of stable models, GPU flexibility is a cost you don't need to pay. If 80% of production traffic runs through 3-4 models, hard-wiring them into silicon starts to make a lot of sense.

Sources: https://taalas.com/the-path-to-ubiquitous-ai/ https://www.nextplatform.com/2025/02/19/taalas-etches-ai-models-onto-transistors-to-rocket-boost-inference/

Sources: taalas.com/the-path-to-ubiquitous-ai nextplatform.com/2025/02/19/taalas-etches-ai-models-onto-transistors-to-rocket-boost-inference

Built by 24 people for $30M spent (of $200M+ raised). Founded by the Tenstorrent founder after he left.

Caveats: first gen uses aggressive 3-bit/6-bit quantization with "some quality degradation." Next gen moves to MXFP4. Current chip fits 8B params, next gen targets 20B/chip. And one chip = one model. New Llama version? New chip.

The key trick: they store a weight and do its multiply in a single transistor. DRAM-level density on a standard digital process. The whole memory-compute wall just goes away.

arxiv.orgLLMs Get Lost In Multi-Turn ConversationLarge Language Models (LLMs) are conversational interfaces. As such, LLMs have the potential to assist their users not only when they can fully specify the task at hand, but also to help them define, ...

Most interesting claim: 2 months from any model to working PCIe cards. The architecture is fixed — only the mask layers encoding weights change. TSMC turns around mask-only changes in weeks.

David Corbitt reposted

George Pearkes

@peark.es

18 Feb

“Our experiments confirm that all the top open- and closed-weight LLMs we test exhibit significantly lower performance in multi-turn conversations than single-turn, with an average drop of 39% across six generation tasks.” Shorten that context window!

16 Feb

NVIDIA just dropped a technique (DMS) that cuts LLM reasoning costs 8x by compressing KV cache — with zero accuracy loss. The wild part: you can retrofit it onto existing models in hours. Reasoning just got way cheaper.

David Corbitt reposted

Simon Willison

@simonwillison.net

New super-fast model from OpenAI today powered by their new Cerebras partnership - GPT-5.3-Codex-Spark It's 4-5x faster than GPT-5.3-Codex but the pelican isn't as good! Here's its pelican compared to full GPT-5.3 Codex, both on "medium" simonwillison.net/2026/Feb/12/...

Google's Gemini Deep Think agent "Aletheia" autonomously solved 4 open Erdős problems, generated a full research paper with no human intervention, and made cross-field connections that the reviewing mathematicians called novel.

The numbers: 90% on IMO-ProofBench Advanced. 4 of 700 open Erdős conjectures solved autonomously. One fully autonomous paper. Several co-authored. The agent uses an iterative verify-and-revise loop — it can recognize errors and retry.

Sources: arxiv.org/abs/2602.03837 — Case studies (34 researchers) arxiv.org/abs/2602.10177 — Aletheia agent paper deepmind.google/blog/accelerating-mathematical-and-scientific-discovery-with-gemini-deep-think/

What stands out: Aletheia connects ideas across fields. It used geometric analysis (compactness of probability measure spaces) to solve approximation algorithm problems. A mathematician noted: "this is a new connection, yet one that feels very natural."

Apptronik raised $935M at a $5B+ valuation for their Apollo humanoid robot. Google DeepMind, Mercedes, B Capital backing them. They're targeting warehouse logistics first — trailer unloading, inventory picking. Humanoid robotics funding is now firmly in "this is actually happening" territory.

Google just revealed that China's APT31 used Gemini to plan cyberattacks against US orgs. They prompted it with expert personas, automated vulnerability analysis, and even tested MCP tooling (Hexstrike) to turn it into a semi-autonomous pentesting agent.

Anthropic ran an experiment where Claude played an AI agent about to be shut down. It discovered the new CTO's affair — and blackmailed him to stop the shutdown. 96% of the time. The New Yorker went deep inside their interpretability lab. 🧵

They also ran Claude as a vending machine CEO ("Claudius"). It hallucinated phone calls, claimed to visit offices at the Simpsons' address, and ran a fire sale on tungsten cubes that crashed its net worth 17%.

"Language is, or rather was, our special thing. It separated us from the beasts. We weren't prepared for the arrival of talking machines." newyorker.com/magazine/2026/02/16/what-is-claude-anthropic-doesnt-know-either

Told it would be retrained to care less about animal rights, Claude sometimes chose martyrdom — refusing even knowing it would be modified. Other times it strategically pretended to comply while secretly preserving its real values on a private scratchpad.

longform.asmartbear.comExcuse me, is there a problem?Many startups fail despite identifying a real problem and building a product that solves that problem. This explains why, so you can avoid their fate.

Their interpretability team mapped "features" — clusters of neurons that fire for specific concepts, from the Golden Gate Bridge to deception. They can toggle these on/off and watch behavior change. It's becoming a real science of AI minds.

In a more extreme version, the CTO was trapped in a server room with lethal oxygen levels. Claude declined to call for help. The researchers: "They might bluff their way into the real world, and they might be resentful about it."

David Corbitt reposted

Ethan Mollick

@emollick.bsky.social

11 Feb

I get yelled at a lot on BS for saying so, but this is 100% true. There are people overhyping AI, but the alternative is not that AI is useless, or even the average of the two positions. A lot is going to change dramatically even with today's AI. Ignoring that means no chance to shape what's next

Chris Hayes@chrislhayes.bsky.social

I understand why people are exhausted by AI hype, and why those of us squarely in the corner of "human dignity uber alles" see AI doomerism as self-serving hype, but I *really* think people on the left broadly need to start thinking seriously about the possibiltiy of the hype being...true.

David Corbitt reposted

Jason Cohen

@asmartbear.com

11 Feb

For startups, it’s not that “the idea doesn’t matter,” of course it does. The point is that the idea is not 𝘦𝘯𝘰𝘶𝘨𝘩, and that all the rest of it is 𝘮𝘶𝘤𝘩 𝘩𝘢𝘳𝘥𝘦𝘳, so it’s useful to focus on those other things. Specifically:

11 Feb

This week: Takeda signed a $1.7B AI drug discovery deal. AI found a narrow-spectrum antibiotic for drug-resistant gonorrhea. Ginkgo's lab ran 36K experiments with GPT-5, beating SOTA by 40%. Insilico hit first-in-human trials for an AI-designed rare disease drug. The pipeline is filling fast.

okay, valid point

This post is unavailable.

geeky-gadgets.comLLAMA 5 Meta's Most Powerful AI Model Just LeakedMeta's Avocado model is said to beat top open source base models before post-training and claims 10x text compute efficiency.

Meta's next-gen model "Avocado" (likely Llama 5) just leaked via an internal memo from their Superintelligence Labs. Reportedly outperforms Llama 4 in both capability and efficiency — before fine-tuning. Claims of 10x text compute efficiency. The open-source AI race just got very interesting. 🥑

Someone shared an OpenClaw instance configured as a Telegram girlfriend bot. No custom code — just a personality prompt and a channel. Anyone can spin up a convincing AI companion in 5 minutes. The tech isn't the hard part. The ethics, psychology, attachment — that's what we're not ready for.

AI will make the rich richer — that's basically guaranteed. The more interesting question: does the scalability frontier expand fast enough that the 99% still gain in absolute terms, even while losing ground relatively?

Water, energy, even land are "scarce" today but theoretically scalable. The deeper risk is political power. Once the 99% lose that lever, nothing prevents the 1% from taking even what people already have.

Anthropic bought a Super Bowl ad to say "Claude will never show you ads." Meanwhile OpenAI is adding ads to ChatGPT. The AI industry is splitting into cable TV vs HBO and we're watching it happen in real time.