AI Reading for Thursday February 13

OpenAI updates roadmap, cancels its hyped o3 near-AGI model by incorporating it in GPT-5 in 'months', and promising a GPT-4.5 non-reasoning Orion successor to GPT-4o in 'weeks'. - TechCrunch

OpenAI in product pivot to an 'oracular system and everything app', will remove complex model picker, automatically send queries to the best model and move to subscription tiers based on intelligence and e.g. allowed rate of deep research queries per day/week/month. - VentureBeat

Pro users definitely want to pick models, choose high v. low effort etc., have access to all the settings in the API though.

DeepSeek and distillation change the environment. If an AI can self-improve and write a better version of itself, then it can also look at all the other AIs and learn from them and incorporate their skills (distillation), so no AI can gain lasting advantage. The singularity is everywhere or nowhere. Maybe a small specialized subject-matter-specific model distilled from a large model makes sense. Maybe OpenAI shouldn’t release the o3 or GPT-5 but only the o3-mini distilled version, since people will use o3 to one-up them at a cheaper price.

---

OpenAI expands their model spec describing how models should act from 10 to 63 pages, details approaches to controversial topics, biases, censorship/free speech - The Verge

---

OpenAI says Musk's proposal to buy the non-profit conflicts with his lawsuit to block changes to OpenAI's structure. - NY Times

Elon Musk says ok fine I will withdraw bid for OpenAI's nonprofit if its board agrees not to restructure. - TechCrunch

A foolish consistency etc., seems to concede bid is just a nuisance tactical ploy.

---

Elon Musk's xAI plans massive hiring spree of thousands of data annotators/tutors to train Grok 3, employees say - Business Insider

Musk says Grok 3 coming soon - Tech.co

---

AI bank analysts reduce human analysts to prompt engineers and proof-readers blubbering, I did an MBA and CFA level 3 for this? - FT

A lot of sell-side research was boilerplate. Who reads this stuff? Have your AI call my AI, do virtual lunch, just tell me what to do.

---

Snowflake adds Claude in Cortex agents platform - VentureBeat

---

Doppelganger dual-interpretation unicode characters and emojis can be used to jailbreak LLMs - Geeky Gadgets

---

Instagram's AI avatar character generator sometimes makes creepy, underage, sexually explicit AI bots - Fast Company

---

UT team Apptronik raises $350m for humanoid robots - TechCrunch

---

Eric Schmidt warns of (the inevitable) AI 9/11 - the Guardian

Also says to focus on open source, cooperate with China on safety so we got that going for us which is nice - FT

---

Google: yeah just be evil now - the Guardian

---

Larry Ellison wants to put all US data in one big Oracle AI system. - The Register

Surely dodgy doge Trumpy Howser interns, Kash Patel and the CIA could do no harm with something like that.

---

AI posters in classrooms complete with hallucinations of the bracha, tracha and diaphagm 🤦 - The Cool Down

---

BuzzFeed's manifesto against Big AI Algorithm - BuzzFeed

Back to our roots doing 7-word listicle headlines optimized for Homo Connectus and not just Homo Algorithmus.

---

5 ways AI powers cybersecurity threats - - PCWorld

Another take on the same subject - CSO Online

---

CalTech AI creates super long mathematical proofs, solves longstanding problems - phys.org

---

Adobe updates video tool to compete with Runway, Chinese AIs, Sora - Reuters

50¢ per 5-second video.

---

AI deepfake of Jewish entertainers against Kanye goes viral - The Jerusalem Post

Scarlett Johansson says, could y’all not deepfake her? Even if you think your intentions are good, putting words in her mouth via an AI likeness is harmful execution - The Hollywood Reporter

---

AI analysis shows Trump has the brain of an 11-year-old - Inquisitr News

---

Follow the latest AI headlines via SkynetAndChill.com on Bluesky