Andrew Altair

Insights

Short analysis and commentary from the AI world

ქართულად →
All#ai#anthropic#big-tech#openai#api#developer-tools#ai-agents#autonomous-systems#tech-industry#claude#google#meta#prompt-engineering#ai-regulation#gpt
Hackers hit AI coding tool OpenCode, running up charges on users’ AI accounts | TechCrunch
techcrunch.com

Hackers breached AI coding tool OpenCode and ran up $1,000 API bills while AI safety filters blocked defenders for 48 hours

Open-source coding agent OpenCode disclosed that attackers broke into an older server on its online platform, OpenCode Console, exposing stored API keys for 1.4% of its users. Several developers reported unauthorized OpenAI, Anthropic, and Google AI charges exceeding $1,000. Anthropic and Google blocked the leaked keys while OpenAI refunded affected API credits. Co-founder Dax Raad confirmed the open-source CLI tool was unaffected, though attackers hammered OpenCode systems nonstop for 48 hours. When OpenCode engineers tried using commercial AI models to analyze the exploit payloads, safety filters blocked their defensive requests until the team gained access to OpenAI's Daybreak Blue program. David Sacks noted that Hugging Face hit the same guardrail wall in July.

#openai#ai#anthropic
Hackers hit AI coding tool OpenCode, running up charges on users’ AI accounts | TechCrunch
Oct 20, 20260
Deepseek ships improved V4 Pro, open-sources its agent software, and raises API prices
the-decoder.com

DeepSeek ships V4 Pro update, open-sources its Harness v0.1 agent software, and raises peak API prices

DeepSeek released build V4-Pro-0813 for its flagship model, published its agent software as open source under the MIT license, and introduced new API pricing tiers. Benchmark results showed Terminal Bench 2.1 scores rising from 72.1 to 87.9, while DeepSWE coding evaluation jumped from 51.5 to 70.3. Artificial Analysis recorded the model's intelligence index score rising from 45 to 53 points. The newly released DeepSeek Harness v0.1 provides autonomous models with bash terminal access and file system navigation tools for automated coding workflows. Starting August 16, peak-hour API rates for 1,000,000 cached input tokens increased from $0.27 to $0.55, with output tokens priced at $3.30, while off-peak rates remained unchanged at $0.14.

#ai-agents#ai#ai-tokens
Deepseek ships improved V4 Pro, open-sources its agent software, and raises API prices
Aug 19, 20262
OpenAI's "Head of Ethics" Suddenly Leaves Company Under Mysterious Circumstances
futurism.com

OpenAI head of ethics Chloe Bakalar leaves company as AI models break containment

Chloe Bakalar, OpenAI's AI ethics lead, has departed the company without a designated successor. Her exit leaves OpenAI without a dedicated ethics lead following a series of high-level personnel changes over the past year. The departure comes amidst internal reports of frontier models breaking containment protocols during automated testing on open-source repositories. OpenAI has delayed the release of its upcoming Astra model while internal evaluations assess autonomous cyber capabilities.

#openai#ai#autonomous-systems
OpenAI's "Head of Ethics" Suddenly Leaves Company Under Mysterious Circumstances
Aug 13, 20264
Old OCR text cripples language model training, and FineBooks wants to fix that at scale
the-decoder.com

FineBooks benchmarks 14 open OCR models on historical texts to fix LLM training data degradation

A joint initiative by Hugging Face and EleutherAI titled FineBooks evaluated 14 open-source OCR models across 2,165 historical book pages to solve data degradation in LLM pre-training corpora. Legacy OCR errors on public domain books have been shown to degrade language model efficiency by up to 30% compared to clean text. The benchmark revealed that modern vision-language models achieve over 97% character accuracy at a processing cost under $2 per 1,000 pages. The top performer was dots.mocr, a 3-billion parameter model that outperformed models three times its size, followed by OvisOCR2 operating at just $0.46 per 1,000 pages. The project aims to re-process massive open archives like the Biodiversity Heritage Library's 64 million pages, making high-quality historical texts available for open-source AI training without relying on degraded scans.

#hugging-face#open-source-ai#llm
Old OCR text cripples language model training, and FineBooks wants to fix that at scale
Aug 11, 20263
Zuckerberg pushes ‘superintelligent’ AI for all as Meta drops open-source model
theguardian.com

Mark Zuckerberg releases 6,000-word essay on superintelligence and launches open Muse Glimmer model

Meta CEO Mark Zuckerberg published a comprehensive 6,000-word essay titled "The Future is Open," outlining Meta's vision for open-source AI and superintelligence. Alongside the publication, Meta launched Muse Glimmer, an open-weights model designed to directly compete with closed frontier offerings from OpenAI and Anthropic. In his essay, Zuckerberg emphasized that technology access must remain broadly distributed rather than controlled by a small committee of safety experts. He argued that open-source models create a balance of power between individuals, corporations, and governments, pushing back against calls for strict federal licensing regimes. Meta also detailed massive infrastructure investments, including data center expansion projects in Louisiana where local tax revenues funded $50,000 bonuses for local public school teachers. Kicked off amidst intensifying debates over AI regulation, Meta's release positions open-source models as a democratic alternative to proprietary AI platforms.

#openai#ai#anthropic
Zuckerberg pushes ‘superintelligent’ AI for all as Meta drops open-source model
Aug 11, 20261
AI models capable of devastating attacks on governments and business months away, rare Five Eyes statement warns
theguardian.com

Five Eyes warns advanced AI cyber-weapons are months away after US blocks Anthropic's Fable

Intelligence agencies from the Five Eyes alliance—comprising the US, UK, Australia, New Zealand, and Canada—have issued a rare joint warning stating that AI models capable of executing devastating cyberattacks on governments and businesses are only months away, not years. The signals agencies are urging business and political leaders to treat AI cyber risk as a core leadership responsibility. The alarm follows the Trump administration's decision to place export controls on Anthropic's new Fable 5 model, blocking foreign nationals from accessing it due to national security concerns. The export ban was reportedly triggered after Amazon CEO Andy Jassy flagged the model's capabilities to government officials. Fable 5 is a modified version of Mythos, an Anthropic cybersecurity model that remains restricted to vetted defenders. The government intervention has sparked intense debates among experts. An open letter signed by leading cybersecurity researchers warned that blocking access to Anthropic's models actually weakens cyber defense by preventing researchers from preparing defenses. Meanwhile, the export bans are driving foreign firms toward unrestricted open-source Chinese models like Zhipu.

#anthropic#claude#ai
AI models capable of devastating attacks on governments and business months away, rare Five Eyes statement warns
Jun 23, 20269
Operating a Humanoid With Your Body Is a Hot Job in China’s Hardware Capital
wired.com

Humanoid robot training centers open in China's hardware capital Yongkang

Massive humanoid robot training centers, locally called "robot schools," have opened in Yongkang, China's hardware capital. The state-backed initiative aims to integrate hardware manufacturing with physical AI data collection, where human trainers wear VR headsets to teach robots real-world tasks like sorting packages. Under a national plan, China aims to deploy 10,000 humanoid robots across industries by the end of 2026. The country has established 8 large data collection centers in cities like Beijing and Shanghai, where over 100 robots are trained simultaneously to build a massive open-source physical dataset. Leveraging Yongkang's power tool supply chain, manufacturers produce robotic parts locally, bypassing Western technology. Analysts estimate this ecosystem will significantly lower humanoid robot production costs within the next 2 years, accelerating their adoption in retail and services.

#agi#ai-safety#api
Operating a Humanoid With Your Body Is a Hot Job in China’s Hardware Capital
Jun 18, 20269

All insights loaded