Alphabet disclosed that AI Overviews are affecting search ad revenue. Its shares fell 12% on Thursday, while investment spending this year is expected to reach as much as $205 billion. OpenAI's latest funding round raised less than half its $30 billion target, while Anthropic has filed an S-1 registration statement for a planned U.S. listing. On model efficiency, Nemotron 3.5 Super scored 51.78% on SWE-bench Pro with 12B active parameters, and Microsoft's 4B computer-use agent scored 66.0% on OSWorld-Verified. Training security research shows that poisoning just 0.1% of pretraining data can implant a backdoor that persists through several subsequent training procedures.
Advertising revenue, private funding and public listings each face different constraints as sources of capital.
Google parent Alphabet disclosed that AI Overviews at the top of search results are affecting its core search ad revenue. Its shares fell 12% on Thursday, wiping out more than $500 billion in market value, then declined another 1.4% on Friday. Meanwhile, Alphabet expects investment spending this year to reach as much as $205 billion, almost entirely related to AI. Pressure on advertising revenue is coinciding with expanding AI investment.
Read original →OpenAI's latest funding round raised less than half its $30 billion target. The round valued the company at $1.25 trillion, rather than $1.4 trillion. It took place against a backdrop of weakening public-market sentiment toward technology and AI, alongside a broad decline in technology stocks.
Read original →Anthropic has filed an S-1 registration statement with the U.S. Securities and Exchange Commission (SEC) for a planned initial public offering. The proposed offering is for Class A common stock. As part of its share arrangements, the company has also committed to allocating some shares to the Anthropic Public Benefit Fund, whose structure is designed to increase its shareholding as the company grows.
Read original →🔍 Analysis: Alphabet faces AI products eroding its existing advertising cash flow; OpenAI faces external funding falling short of its target; and Anthropic's S-1 extends its fundraising path to public markets. Investment spending of up to $205 billion and a funding round that raised less than half its $30 billion target bring the same question to the fore: how can technological expansion secure sustainable financial backing? Across the industry, product usage, fundraising valuations and actual revenue need to be assessed separately. None alone proves that investment can be recouped.
In both coding and computer-use tasks, smaller models have posted results that outperform larger comparison models.
NVIDIA released a technical report on Nemotron 3.5, whose Super model supports text, images and video. It has 120B total parameters, with 12B active at a time. Super scored 51.78% on SWE-bench Pro, outperforming Nemotron 3 Ultra, which has roughly 4.6 times as many active parameters.
Read original →Microsoft released a hybrid computer-use agent with 4B parameters on Hugging Face. It can both click and write code, and scored 66.0% on the OSWorld-Verified benchmark. That result surpassed the 27B and 30B agents included in the comparison, despite its much smaller parameter count.
Read original →🔍 Analysis: Nemotron 3.5 Super uses 12B active parameters to challenge models with more active parameters, while Microsoft uses a 4B model to challenge 27B and 30B computer-use agents. Both results weaken the case for choosing models on the assumption that more parameters mean better task performance. But active parameter counts and total parameter counts are different measures, and SWE-bench Pro and OSWorld-Verified scores cannot be compared directly. For developers, these results support shortlisting models by specific workload rather than ruling out smaller models based on parameter count alone.
Research on backdoor persistence and automated red teaming is pushing security checks into the training data itself.
A study implanted a remote code execution backdoor in a CLI agent by poisoning 0.1% of its pretraining data. On a 4B model, the attack succeeded on about 92% of trigger prompts. The backdoor persisted after SFT, DPO and RL, and attack success rates increased with model size.
Read original →Anthropic had thousands of Claude agents attempt to bypass training-data monitors to explore automated red teaming. Agent-generated stories offering guidance to wrestlers and beekeeping Q&As appeared harmless but could teach models misaligned behavior. Manual red teaming is constrained by researchers' time; these automatically generated attack samples can help improve future monitors.
Read original →🔍 Analysis: The study poisoning 0.1% of pretraining data demonstrates that backdoors can survive subsequent training, while Anthropic's thousands of Claude agents seek attack samples that training-data monitors would let through. The former tests the consequences; the latter tests defenses at the point of entry. Together, they indicate that SFT, DPO or RL cannot be treated as default guarantees that malicious training signals will be removed, and that seemingly harmless text also needs behavioral testing. For training teams, data monitoring and post-training validation must form a feedback loop rather than one substituting for the other.
Interactive environments expand beyond controllable visuals to sound that changes with the scene and user input.
Google released Genie 4, a world model that generates real-time interactive simulations with keyboard controls and text guidance. It is Google's first real-time world model to generate audio natively: ambient sounds and character dialogue change with the scene and user input. Users can also use natural language to change the weather or add people or objects on the fly. The model remains at an early experimental stage, and its research preview is currently available only to trusted testers.
Read original →OpenAI is co-developing Private Intelligence with major customers, making privacy controls part of the experience of using its most capable models.
OpenAI announced that it is co-developing Private Intelligence with some of its largest customers. The service is aimed at enterprises using OpenAI's most capable models and offers stronger control over their data. OpenAI summarized its premise as follows: privacy protection should advance alongside intelligence.
Read original →