Thibault says users can expect 2 banked resets over the next two days That means we'll soon have a pretty substantial stockpile of tokens for... GPT-6.1 ASTRA, OF COURSE! Or something equally massive. 🤔 https://t.co/L5rvPsSQjW

A developer reverse-engineered his $3,000 Samsung TV after suspecting its targeted ads weren't a coincidence. He uncovered extensive tracking and telemetry, including automatic content recognition, audio monitoring, and user profiling. OpenAI then banned his account for alleged “cybersecurity abuse.” He says he only analyzed local binaries, never used OpenAI to find or exploit vulnerabilities, and lost around $500 in accumulated usage credits. His response? Buy three more computers to expand his local AI cluster. “It's much nicer to own your own AI.”

OpenAI has unveiled four major scientific results achieved by its own experimental AI model, which remains unavailable to the public. 1. Mathematics: Proved the Kakeya conjecture in all dimensions, advancing beyond previous breakthroughs in 2D (2019) and 3D (2025). 2. Computer science: Disproved the Exponential Time Hypothesis (ETH), challenging assumptions underpinning hundreds of studies and suggesting some hard computational problems may be solvable much faster. 3. Physics: Mathematically proved the consistency of quantum chromodynamics (QCD) and a unified version of the Standard Model. 4. Yang–Mills theory: Proved the theory's existence, solving part of a Millennium Prize Problem worth $1 million. The proofs in mathematics, computer science and Yang–Mills theory passed independent computer verification. Leading experts, including researchers behind the disproved hypothesis and a Nobel laureate, are now examining the results. OpenAI says it will not seek scientific credit for these findings. 👏

OpenAI has released GPT-6.1 Terra. It’s slightly weaker than Sol but cheaper, with particularly strong computer-use capabilities. The model is rolling out to all paying users. API pricing: $1 per million input tokens and $5 per million output tokens. https://t.co/BPnVr8YUsF

Galaxy Research’s VC Vital Signs report (Oct 23) highlights a crypto funding paradox in Q3 2026: startup valuations hit a record high, while median deal size fell 40%. Three key takeaways: 1. Early-stage deals remain active. Pre-seed rounds accounted for ~25% of deals, versus 24% for late-stage companies. These figures reflect deal counts, not capital invested, so they don't mean pre-seed attracted more funding. 2. Valuations rose as deal sizes shrank. VC-backed crypto company valuations reached ~$90M, the highest in the tracked dataset. Meanwhile, median deal size dropped from $5M in Q2 to $3M in Q3, down 40%. These metrics measure different things: valuation reflects a company's agreed value during fundraising, while deal size measures the capital raised. The divergence may indicate that funds still accept high valuations but are more cautious about how much they deploy. Changes in the composition of companies surveyed could also affect the data. 3. VC funding isn't token market inflows. Most deals involve company equity or private investment agreements, with different valuations, allocations and unlock schedules. Rising startup valuations don't automatically translate into higher token FDVs, nor do they prove a new altcoin bull cycle has begun. If this trend continues, crypto startups may still raise at relatively high valuations, but with smaller rounds and greater pressure to use capital efficiently. Source: Galaxy Research, VC Vital Signs, Oct 23, 2026.

ZCode is giving away 500 million free tokens for GLM-5.3 and GLM-5.3-Flash to celebrate China's Programmer's Day. - 200M tokens for GLM-5.3 - 300M tokens for GLM-5.3-Flash Even free-tier users can claim a share of the 200,000 available packages, distributed on a first-come, first-served basis. The promotion runs from October 23–26. Tokens can only be used within ZCode, not via API. If you have ZCode installed, update it and check. 😏

fal has pushed FLUX 3 close to real-time video generation. Working with BFL, fal fine-tuned the model and optimized inference to create FLUX 3 Max, which generates videos with audio from text or images. fal's benchmark: 10 seconds of 720p video in 12.45 seconds vs. 277.8 seconds for Seedance 2.5 — roughly 22x faster. On Artificial Analysis's I2V Arena, it ranks among the leaders alongside Seedance 2.5 and MiniMax H3. Try it on fal for $0.0425/sec until Nov 7 ($0.425 per 10-second clip). After the promo, that clip costs $1.70.

Every Claude Code user should install /stack. It scans EVERYTHING available in Claude Code: - Skills, subagents, commands, hooks, MCP servers, plugins, settings, and output styles. - Config layers: Enterprise, User, Project, Plugin, and CLI flags. An interactive dashboard shows what's active, disabled, or overridden, helping explain token usage and Claude's behavior. It even reveals built-in commands, skills, and subagents missing from the official docs. Install: npx claude-code-templates@latest --mod ui/stack-viewer

OpenAI found that usage limits were incorrectly set for around 40,000 Codex accounts, causing credits to be consumed faster than expected. The issue is fixed, and affected users will receive $500–$1,000 worth of free credits. One user has already reported receiving over 12,000 credits. Check Settings → Usage & billing → Credits in Codex, or check your email. 😏

A developer tested Astra and DeepSeek V4.1-1017 Flash with the same prompt and maximum reasoning effort. The results surprised me. - Astra: 20 minutes, $7. Better attention to detail and a more original result than GPT models typically produce. - DeepSeek: 1 hour 40 minutes, $0.90. An absolutely stunning environment, but the koi fish didn't turn out so well.