Marcus Ting

Founder of Anthera - Spatial Intelligence and Artificial intelligence company - Intelligent Systems for Human Capability - Australia - Guitar enthusiast

[Marcus Ting @marcusting]

Claude Opus 5 is out, same price as Opus 4.8. Agentic terminal coding jumps to 43% (Opus 4.8: 21%, Fable 5: 33%), and it beats Fable 5 on computer use and business workflows for less. Also notably better at verifying its own work. #AIAgents #EnterpriseAI

[Marcus Ting @marcusting]

OpenAI's CFO wants AI judged on cost per successful task, not cost per token. Fair test, though it flatters their own cheapest model. PwC says 56% of CEOs see no financial benefit from AI yet, and most are spending with no way to measure it. #EnterpriseAI #AgenticAI

[Marcus Ting @marcusting]

Gemma 4 12B, out 3 June: a 12B open model that runs fully on a 16GB laptop, no cloud, no API key. It writes and runs code locally and serves an OpenAI-compatible endpoint your tools can point at. For data that can't leave the building, the on-prem line moved.

#EnterpriseAI

[Marcus Ting @marcusting]

Claude Fable 5 is live today. A Mythos-class model, a tier above Opus, topping nearly every benchmark, at $10 per million input tokens.

When the frontier goes public, the model stops being your edge. Governance and smart routing are.

#EnterpriseAI #AgenticAI

[Marcus Ting @marcusting]

Australia's Privacy Act automated decisions clause takes effect December 2026. Eight months to be able to explain why an AI made a decision about a person. Most enterprises cannot answer that question today.

#AIGovernance #PrivacyAct

[Marcus Ting @marcusting]

DeepSeek V4 ships this week. Performance near frontier. Cost roughly one-sixth of GPT-5.5. The arbitrage gap between models is now wide enough that workload routing matters as much as model selection.

#EnterpriseAI #AIAutomation

[Marcus Ting @marcusting]

Australia skipped the AI Act. The Dec 2025 plan leans on existing law. But the Privacy Act amendment is passed: from 10 Dec 2026 you must disclose the data behind automated decisions. No AI Act doesn't mean no obligations. #EnterpriseAI #AIGovernance

[Marcus Ting @marcusting]

Opus 4.8 shipped six weeks after 4.7. For production agents the number that earns trust is reliability. This one is half as likely to claim work it never finished, and 4x less likely to wave bad code through. That is what makes autonomy deployable.

#AgenticAI #EnterpriseAI

[Marcus Ting @marcusting]

Google committed $40B to Anthropic. $10B immediate, $30B performance-based. Plus Amazon's $25B. The hyperscalers are buying compute futures as much as equity. A multi-polar AI market is better for enterprise procurement than any single dominant vendor.
#EnterpriseAI

[Marcus Ting @marcusting]

Meta Superintelligence Labs released Muse Spark on April 8. No open weights. After three years of Llama anchoring the open-weight thesis, Meta's most capable model is closed. Open weights are a per-release choice now.

#EnterpriseAI #AIStrategy

[Marcus Ting @marcusting]

The NSA is using Anthropic's Mythos despite a procurement blacklist. The agency that runs signals intelligence chose to bypass its own rules to access AI capability. Every enterprise has a quieter version of this story right now.

#EnterpriseAI #AIGovernance

[Marcus Ting @marcusting]

The leaked GPT-5.5 Codex system prompt contains the line "Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant." Constraint shape is deployment shape.

#AIAgents

[Marcus Ting @marcusting]

Model upgrades aren't just benchmark improvements. The Opus 4.6 to 4.7 system prompt changes show how default model behaviour shifts between versions. Worth a read if you deploy AI at scale. #EnterpriseAI #AgenticAI

[Marcus Ting @marcusting]

#SpatialComputing #EnterpriseXR #EnterpriseXR #EnterpriseXR #EnterpriseXR #EnterpriseXR #EnterpriseXR #EnterpriseXR #EnterpriseXR #EnterpriseXR #EnterpriseXR

[Marcus Ting @marcusting]

Anthropic's agent experiment: Opus got $65 for a broken bike. Haiku got $38. Same instructions. Users on the weaker model had no idea. Model quality is a commercial variable. #EnterpriseAI #AIAgents

[Marcus Ting @marcusting]

#SpatialComputing #AIAgents #AIAgents #AIAgents #AIAgents #AIAgents #AIAgents #AIAgents

[Marcus Ting @marcusting]

#AIAgents #AgenticAI #AgenticAI #AgenticAI #AgenticAI #AgenticAI #AgenticAI #AgenticAI #AgenticAI #AgenticAI #AgenticAI

[Marcus Ting @marcusting]

Before any enterprise AI contract in mining, healthcare, or government: source citation, confidence score, model attribution. If the vendor can't show all three on a live query, the tool isn't field-ready. #EnterpriseAI #AIAgents

[Marcus Ting @marcusting]

Claude Opus 4.7 shipped task budgets. The model sees a countdown of tokens and wraps its loop gracefully.

Agentic deployments fail when the agent runs too long and burns too much budget. Control primitives are the real enterprise story.

#AIAgents #AgenticAI

[Omar (mainnet arc) @acceleratooooor]

Here's how to triage:

1. Go to https://t.co/bfAEiJ6emr

2. Security → Access and data control → API controls → App access control → Manage Third-Party App Access

3. Search for client ID:
110671459871-30f1spbu0hptbs60cb4vsmv79i7bbvqj

if found → revoke / block