Tech • AI • Robotics • Game

VIDEO
ENFR
TodayPlay for freeFury BloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

OpenAI Decisions API, GPT Next leaks, GLM 5.3 exploits

AITuesday, October 6, 2026· 16 videos

Briefing

Audio player
0:00 / 0:00

OpenAI debuts Decisions API

OpenAI introduced the Decisions API, a low-latency interface built on GPT-6 Luna for routing text or image inputs into predefined actions. The product is aimed at operational tasks such as support triage, lead qualification and agent handoffs where open-ended generation is unnecessary. OpenAI says responses can arrive in under 100 milliseconds, roughly an order of magnitude faster than broader chat-style calls. The launch underscores a wider shift toward specialized inference products that compete more directly with classic classifiers than with chatbots.

GPT Next leaks amid ship sprint

A previously unseen GPT Next model surfaced in early testing as OpenAI pushes through a stated 28-day release cadence. Early impressions suggest better performance on web development and 3D scene generation, with unusual attention on potentially lower pricing rather than a dramatic capability jump. Separately, GPT-6 Astra and GPT 6.1 Sonnet were reportedly made about 50% faster, lifting target generation speed toward 50 tokens per second. Together, the signals point to a strategy centered on product tempo, efficiency and cost pressure as much as on frontier leaps.

GLM 5.3 shrinks exploit gap

New analysis of Zhipu AI's GLM 5.3 suggests advanced offensive cyber capability is becoming cheaper and more accessible. On a benchmark for writing Chrome exploits, the open-weights model reportedly achieved a full exploit in about 12% of attempts, close to 14% for Anthropic's Mythos, while the prior generation produced none. Researchers also described a workflow in which the model found previously unknown flaws and chained them into an attack capable of stealing a private key from a victim machine. The practical implication is that defenders may need to prioritize speed, telemetry and automated response over assumptions that high-end exploit development will remain scarce.

Codex CLI expands terminal workflow

Codex CLI added a full-screen terminal interface, voice-triggered multitasking, remote phone control and richer in-terminal rendering. The update also introduces task views for multi-agent monitoring, follow-up instructions for running jobs and conversation forking through worktrees. Native rendering for Mermaid diagrams and LaTeX reduces the need to leave the command line for visualization and documentation. The package reflects the industry's push to turn the terminal into a full orchestration surface for AI-assisted software work.

Type Safe pitches Jeev engine

Type Safe, a San Francisco startup founded by former InstructGPT contributor Diogo Almeida, launched Jeev after raising $40 million in seed funding. Instead of generating long text, Jeev selects among bounded options and returns a probability score, positioning it as a fast decision engine for tightly scoped tasks. The architecture revives an encoder-style approach closer to flexible classification than to decoder-heavy chat systems. Its bet is that many production automations need cheap, reliable routing and labeling more than they need fluent language generation.

Ajax revives uncensored open models

A new model called Ajax was released for the open-source agent platform Odysius, combining Qwen 3.5 9B with task-specific fine-tuning and reinforcement learning. The project is tied to Felix Kjellberg and deliberately removes refusal behavior, producing a model that answers prompts mainstream systems reject. Developers said uncensoring relied on parameter-removal techniques associated with projects such as Heretic. The release highlights the widening split between commercial safety guardrails and open-weight communities optimizing for local control and unrestricted behavior.

AI schools redraw San Francisco elite track

Private education groups in San Francisco are accelerating AI-centered school models that compress core academics into roughly two hours a day of adaptive software. Alpha School is expanding with tuition between $40,000 and $75,000 and ambitions for about 50 U.S. campuses, while Andreessen Horowitz has committed $35 million to launch The Academy in September 2027. Proponents argue the model frees time for coding, entrepreneurship and public speaking. Critics see a risk of deepening inequality by building a premium pipeline that blends software tutoring with venture-style networks and coaching.

Humanoid race splits three ways

Three robotics efforts illustrated diverging commercialization bets in humanoids. Clone Robotics showed Torso 3 using a musculoskeletal design with artificial muscles and fluid control, even as it said Torso 4 is already replacing parts of that platform after a six-month redesign. Robo Party promoted an open-source biped focused on resilience, while Agibot said it had placed retail service robots in 100 stores in China. The contrast captures a market still undecided between biomimetic motion, open developer ecosystems and narrow service deployments.

Videos covered

Previous briefings · AI