In partnership with
Ai2 released AstaBrief on 2 October. The 8-billion-parameter model drafts cited reports from research questions and retrieved literature excerpts. A general assistant may be asked to search the web. AstaBrief instead receives selected evidence with the query. That distinction changes what the model can answer on its own. Report quality depends on the writing model and retrieved evidence. It also depends on whether each sentence represents its sources accurately.
200+ Proven Ways to Make Money With AI in 2026
The next wave of millionaires will be people who figured out how to make AI work for them.
The window to get ahead is still open. But not for long.
Here are 200+ proven ways to make money with AI in 2026.
Sign up for Superhuman AI, the free daily newsletter read by 1M+ professionals, and get instant access to all 200+ ways to profit from AI this year.
Ai2 built the model from Qwen3-8B, then used supervised fine-tuning and direct preference optimization. The institute says its training examples came from a prior multi-step literature pipeline. It first gathered relevant papers and organized them by section. Then it used outputs from larger models, including Claude and OpenAI systems. Ai2 also released the model weights, training data, and an example workflow. Researchers can inspect more of the system than a hosted interface normally exposes.
AstaBrief’s main design choice is single-pass report generation. It produces a report from a question and retrieved text in one pass. There are no separate summarization, clustering, or section-writing stages. In Ai2’s comparison, Fast averaged 51.1 seconds per report. Its Thinking pipeline averaged 178.5 seconds. Both are company-reported timings for Ai2’s setup. They are not universal comparisons across models or hardware.
The shorter pipeline also changes where review can happen. Multi-stage systems may expose outlines or intermediate summaries. They can show section-level results before the final answer. A single-pass output offers fewer natural checkpoints. A reviewer may miss an absent paper, weak section, or claim that exceeds its source. Faster generation does not remove the need to inspect retrieval. Reviewers must still check citations and the finished text.
Build or buy your support agent? Both skip the number.
Buying or building your customer service agent both carry a cost the pitch skips: maintenance if you build, flexibility if you buy. Running Agents in Customer Work is four conversations on agentic AI in customer ops, this one on build vs. buy. Register now, four Tuesdays, 10 a.m. PT.
Hugging Face’s editorial analysis separates three questions. It notes that evaluations can collapse them into one score. Coverage asks whether the answer addresses the research question. Attribution asks whether claims point to relevant sources. Entailment asks whether the sources support the wording. It also checks the population and strength of each claim. A citation can be relevant while the sentence overstates what the paper found. A high citation count alone cannot establish accuracy.
Ai2 also puts a time boundary on its results. Ai2 says much of its training and evaluation occurred in 2025. It did not rerun the full evaluation against current frontier models. The results are evidence about the training and design choices Ai2 tested. They do not show that AstaBrief outperforms today’s leading systems. Its model card describes a research-reporting task, not every scientific field or consequential decision.
Open weights allow local processing. They do not, by themselves, guarantee privacy or reliability. Institutions must review document storage, access controls, logs, and retrieval quality. They should identify any external services used. Local hardware does not prevent a model from repeating misleading text. That text may come from its supplied material. Readers need to know which papers entered the model’s context. They also need to know whether conflicting evidence was excluded.
For research teams, the release is best read as a design case. It is not a final verdict on automated literature review. A useful local test would include conflicting studies and thin evidence. It should also vary populations and test unsupported assumptions. Reviewers could mark missing coverage, weak attribution, and overstatement separately. That makes errors easier to locate than one broad quality score. For today’s research note, a citation is useful when it lets readers test the exact claim against its source.
How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads
For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.
LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.
The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.
Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV ads.
The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.
