Skip to content
Trending

Worth knowing

Ideas that outlast the news cycle. There are no dates on this page: each idea was true when we checked it and will still be true next quarter. Not everything that spreads on social media is true, so every claim is checked against its primary source and explained in our own words.

14 ideas · explained, then checked

Each card says what is going around, what is actually true, the one line to remember and what it means for a product. Open the checks to see the source behind every claim.

58 claims · 27 hold · 25 partly · 6 wrong

Concept

9 of 10 widely used benchmarks score 'I do not know' as zero, so language models learn to bluff.

The usual explanation is that a language model invents facts because it only predicts the next word, and that connecting it to your own documents makes the problem go away. A fuller version says training and testing reward a confident guess over an honest 'I do not know', and that the fixes with evidence are grounded answers with citations, permission to abstain, and scoring that punishes confident errors.

  • 9 of 10benchmarks with no credit for abstaining
  • 26% vs 75%wrong answers on SimpleQA, declining vs guessing model
  • 74.5%citations that support their sentence
  1. Models state false things confidently because training and testing reward a guess over 'I do not know'.Kalai and colleagues at OpenAI and Georgia Tech show the errors start in pretraining: facts with no pattern, such as birthdays, can only be memorised, so if 20 percent of them appear once in the training text, expect the base model to get at least 20 percent wrong, even with error-free data. The errors then survive because of grading. Under binary scoring, one point for a right answer and nothing for a wrong one or an abstention, declining is never the best move. Of the 10 widely used benchmarks the authors reviewed, 9 score this way. Only WildBench gives abstention partial credit, and the authors note its rubric can still rate 'I do not know' below a flawed answer.Why Language Models Hallucinate (Observation 1, Table 2)
  2. Connecting the model to your documents removes the problem.Retrieval raised the share of correct answers in a study by researchers at Google, UC San Diego and Duke, but it also made models less willing to decline. Claude 3.5 Sonnet abstained on 84.1 percent of questions without retrieved text and on 52 percent with it; Gemini 1.5 Pro went from 100 percent to 18.6 percent; GPT-4o barely moved, from 34.4 to 31.2. On the Musique question set, 55.4 percent of questions came with passages that did not hold enough to answer, and the larger models often answered wrongly instead of declining when the passages fell short. A retrieval miss does not look like an error in production; it looks like a fluent answer.Sufficient Context: A New Lens on Retrieval Augmented Generation Systems (Section 4.2, Figure 1)
  3. Answers that cite their sources can be trusted.Grounding has evidence behind it: in the original RAG paper by Lewis and colleagues, human raters comparing 452 pairs of generated quiz questions judged the retrieval model more factual in 42.7 percent of pairs and the plain model in 7.1 percent. A citation is still a claim to be checked. A Stanford audit of four generative search engines, Bing Chat and Perplexity among them, found 51.5 percent of generated sentences fully supported by their citations, and 74.5 percent of citations supporting the sentence they were attached to. Roughly 1 citation in 4 did not back its sentence.Evaluating Verifiability in Generative Search Engines
  4. Letting a model decline cuts false answers without costing much accuracy.OpenAI's GPT-5 system card gives SimpleQA scores without web access for two small reasoning models. o4-mini: 24 percent correct, 75 percent wrong. gpt-5-thinking-mini: 22 percent correct, 26 percent wrong. The card prints no declined share; if every question is answered right, answered wrong or declined, the remainders are 1 percent and 52 percent. Two points of accuracy went with a wrong-answer rate about a third the size (26 against 75). This is one vendor comparing two models that differ in more than their willingness to decline, and the trade is not always this cheap: in the Sufficient Context study, fine-tuning a Mistral model to decline more often cost it correct answers. On an accuracy-only leaderboard the first model still ranks higher, which is the incentive problem in the first check.GPT-5 System Card (Table 8, SimpleQA)
  5. Scoring that penalises confident errors fixes the incentive.The arithmetic holds; the paper offers it as a proposal, not a measured result. State a confidence target t in the instructions and charge t/(1−t) points per wrong answer, so a guess pays only when the model is more than t sure. At t = 0.9 a wrong answer costs 9 points, and an answer that is 85 percent likely to be right scores 0.85 − 0.15 × 9 = −0.5 against 0 for declining. (The paper prints a penalty of 2 for t = 0.75; its own formula gives 3.) The authors argue the rule must go into the main benchmarks, because adding separate hallucination tests leaves the incentive in place.Why Language Models Hallucinate (Section 4.2, explicit confidence targets)

the line to remember

Models bluff because bluffing scores well; score a wrong answer below 'I do not know', ground each claim in a passage, and check the passage supports it.

For your product

Accuracy alone is the wrong number to buy on. Ask a vendor for three figures measured on your own questions: how often the system is right, how often it is wrong, and how often it declines. In acceptance tests, score a wrong answer as worse than a declined one, include questions your documents cannot answer so you see what happens when search comes back thin, and sample the cited passages to confirm they support the sentence. A system that sometimes declines and is rarely wrong is usually worth more than one that always answers.

Sources: Why Language Models Hallucinate (Kalai et al., OpenAI and Georgia Tech) · Sufficient Context: A New Lens on Retrieval Augmented Generation Systems (Joren et al., UC San Diego, Duke University and Google) · Evaluating Verifiability in Generative Search Engines (Liu, Zhang and Liang, Stanford University, Findings of EMNLP) · Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., Facebook AI Research, UCL and NYU, NeurIPS) · GPT-5 System Card (OpenAI)

Concept

5 times the precision at near-equal recall in one open test: 200-token chunks against an 800-token default.

Guides to retrieval-augmented generation tend to hand out a recipe for splitting documents: choose a chunk size such as 512 or 800 tokens, add 10 to 20 percent overlap, and split on structure or meaning instead of fixed lengths. Contextual retrieval and late chunking are then presented as the fix for chunks that lose their surrounding context.

  • precision, recursive 200-token chunks against the 800/400 default (Chroma)
  • 5.7% → 1.9%missed top-20 retrievals with contextual retrieval, BM25 and a reranker (Anthropic)
  • +1.8nDCG at 10 points from late chunking, 52.2 to 54.0

the line to remember

Chunking is a setting you measure on your own documents and questions, because the best splitter, size and overlap changed with the data or the embedding model in each study cited here.

For your product

Ask a vendor which chunk sizes and splitters they tested on your documents, and for the recall and precision of each. If the answer is that they use the default, nobody measured. Start with the cheap option, a paragraph-aware splitter at a few hundred tokens, build a test set of real questions paired with the passages that answer them, and pay for per-chunk context or embedding-based chunkers only when that test shows the gain.

Concept

Flat chunks throw away the table of contents. Keeping it was worth 5.7 points in IBM's STAIR test, not 23.1.

A retrieval method credited to IBM researchers is going around: instead of cutting documents into equal chunks and embedding each one, let the model use the document's own structure, its table of contents and section hierarchy, to find the right section. It is usually presented as a replacement for chunk-and-embed retrieval.

  • 5.7 ptsSTAIR over the same tuned model with no contents page
  • 13.9 ptsSTAIR over an untuned dense retriever, right section ranked first
  • 2.0 ptsRAPTOR's tree over a dense retriever, same reader, QuALITY

the line to remember

Headings are retrieval signal the author already wrote, so keep them through parsing and let search use them alongside embeddings, not instead of them.

For your product

Before paying for a bigger embedding model, check what your ingestion does to headings. Manuals, policies, contracts and textbooks arrive with a contents page; if the parser flattens it, every chunk loses its address. The cheap version needs no training: chunk along the document's own elements and store the heading path with each chunk, which the chunkers in Docling, an open-source parser started at IBM Research, do for you. None of the papers above measures that cheap version, so test it on your own questions. The trained versions (STAIR, RDR2) cost a fine-tuning run each, and when the comparison is fair the published gains are single digits: 5.7, 3.4 and 2.0 points, not 20. None of this helps a corpus with no structure, such as chat logs or email. Ask any vendor two things: does the pipeline keep section headings, and was the benchmark run on documents like yours with baselines tuned the same way.

Concept

A router decides once and steps aside. A supervisor keeps deciding until the job is done.

Multi-agent diagrams often draw a router and a supervisor as the same box: something at the top that picks which agent gets the work. The usual claim is that the two words are interchangeable because both choose an agent.

  • 4 vs 3model calls for one simple request, supervisor vs router
  • 90.2%supervised team over a single agent, Anthropic internal eval
  • 15×tokens of a multi-agent run compared with a chat

the line to remember

A router chooses once and leaves; a supervisor chooses, reads the result and chooses again until the work is done.

For your product

When a vendor draws one box above several agents, ask whether that box decides once or stays in charge. If requests fall into clear categories that one specialist can finish, a router is cheaper, faster and easier to test, because you can score it like any classifier. If a request needs several specialists and the next step depends on what the last one found, you need a supervisor, and a budget for its extra calls and for tracing its decisions. Amazon Bedrock makes the choice a single setting, SUPERVISOR or SUPERVISOR_ROUTER, allows up to 10 collaborator agents per supervisor and notes that the routing mode reduces latency.

Concept

The OAuth standard lists 5 problems with giving software your password. All 5 apply to AI agents.

Common advice for connecting an AI agent or an MCP server to email and other accounts: never give it the real account password. Use OAuth with narrow scopes, fall back to an app password only where OAuth is missing, keep every credential revocable, and for remote MCP servers rely on the authorisation model in the MCP specification.

  • 5problems with password sharing, RFC 6749
  • 14Gmail API scopes to choose from
  • 8 of 14Gmail scopes Google marks restricted

the line to remember

An agent should hold a token that names what it may do, expires, and can be switched off on its own; a password does none of the three.

For your product

Before an agent touches a mailbox, a drive or a CRM, ask the vendor three things: which scopes it requests and why each is needed, where the tokens are stored, and how you revoke one agent without locking out your staff. Treat any connect screen that asks for the account password in the product's own form as a failed review. For MCP servers, ask whether the server follows the specification's authorisation rules or reads a long-lived secret from a config file, because the protocol allows both.

Concept

About 100 tokens per installed skill: the rest of the folder stays on disk until a task matches.

Agent skills are described as a way to teach an agent a job once: a folder with a SKILL.md file holding instructions, scripts and resources, which the agent reads only when a task needs it. Common add-ons to that story are that a skill is just a saved prompt, that skills make tools and MCP servers unnecessary, and that a skill is only text, so it can be installed from anywhere.

  • ~100 tokensper installed skill, at rest
  • <5k tokensSKILL.md body, loaded on match
  • 3 levelsmetadata, instructions, resources

the line to remember

A skill is a folder of know-how that costs about 100 tokens until a task matches it, so write it as you would brief a colleague joining the team and vet it as you would any program you install.

For your product

The know-how that makes an agent useful in your company (how a report is laid out, which checks a refund needs, how a release is cut) can live in version-controlled folders that your team reviews like code. Because the format is an open standard, the same folder can be read by agent products from several vendors, which lowers the cost of switching, though bundled scripts still depend on what each environment allows (network access, installed packages). Put skills under the same controls as software: a named owner, review before install, and no unvetted downloads on machines that hold customer data.

Concept

33 percent less error than seasonal naive on 97 public tasks it never trained on: pretrained forecasters.

Forecasting is said to have its own foundation models: one pretrained model that predicts sales, traffic or sensor readings it has never seen, with no training on that data, and matches models built for each dataset. Google Research's TimesFM is the usual example, described as open weights, able to read about 16,000 past points, and first on every public benchmark.

  • 33%lower point error than seasonal naive, zero-shot (GIFT-Eval)
  • 330Mparameters, version 3.0 (200M in 2.5)
  • 15,360past points read at most
  • 1T+time points in pretraining, up from 100B in the first model

the line to remember

A pretrained forecaster is a strong baseline you get before building anything, so backtest it zero-shot on your own history beside seasonal naive, and make any custom model beat both before you pay for it.

For your product

A first forecast no longer needs a data science project per product line. The same model family runs inside BigQuery as AI.FORECAST with no model to train. Check three things before relying on it: the licence (3.0 weights are non-commercial, 2.5 is Apache-2.0, a managed cloud service has its own terms), a backtest on your own history against seasonal naive and your current method, and whether your real drivers such as price, promotions and holidays can be passed in as covariates. If the backtest will inform a business decision, run it on the Apache-2.0 version or the managed service, because the non-commercial licence rules that use out. Public benchmarks rank models on public data; your data decides.

Concept

Elastic weight consolidation puts a spring on each weight, stiffest where an old task would break.

Elastic weight consolidation (EWC) is usually presented as the cure for catastrophic forgetting, the way a neural network loses an old task while it is trained on the next one. The common account: after the first task, score every weight by its Fisher information, then penalise changes to the high-scoring weights during later training. It is often said to let one network keep learning task after task with nothing lost.

  • 3numbers stored per weight under EWC: value, anchor, importance
  • 20.01%EWC on split MNIST when it must tell tasks apart
  • 90.79%generative replay on the same test

the line to remember

EWC charges the network for moving the weights an old task depends on, which slows forgetting without ending it, and replay still beats it once the network has to tell the tasks apart by itself.

For your product

When a supplier fine-tunes a model on your data, ask what it lost as well as what it gained. A study of language models from 1 to 7 billion parameters found forgetting was the general pattern when instruction tasks were tuned in one after another, and it grew worse with model size in that range, so insist on before-and-after scores for the general skills you depend on. The usual protections are a penalty such as EWC (two extra numbers held per weight, plus a search for the right strength), mixing some earlier data back in, or training a small adapter such as LoRA, which one comparison on code and maths found forgets less than full fine-tuning and also learns less.

Sources: Overcoming catastrophic forgetting in neural networks (Kirkpatrick et al., DeepMind and Imperial College London (PNAS)) · On Quadratic Penalties in Elastic Weight Consolidation (Ferenc Huszár) · Three scenarios for continual learning (van de Ven and Tolias, Baylor College of Medicine) · Full-Parameter Continual Pretraining of Gemma2: Insights into Fluency and Domain Knowledge (Šliogeris et al., Neurotechnology) · An Empirical Study of Catastrophic Forgetting in Large Language Models During Continual Fine-tuning (Luo et al., Tencent WeChat AI and Westlake University) · LoRA Learns Less and Forgets Less (Biderman et al., Columbia University and Databricks Mosaic Research)

Interview question

10 users can crash a 48 GB GPU running a 13B model. The maths says 60 GB.

An interview-style post: a 13-billion-parameter model is deployed on a 48 GB GPU, only 10 users are on it, and the server still dies with CUDA out of memory. The caption blames the KV cache, uncontrolled concurrency and memory fragmentation, and prescribes continuous batching (vLLM, TensorRT-LLM, TGI), capping tokens and context, keeping 10 to 15 percent of memory free, and queueing with backpressure.

  • 26 GB13B weights, fp16
  • 0.8 MBKV per token, 13B
  • 60 GB10 users × 4k tokens

the line to remember

Serving memory = weights + (tokens × concurrent users × per-token KV cost). Size the card for the cache, not the model.

For your product

If a vendor quotes you a GPU from the model size alone, the quote is wrong. Ask for the per-token KV cost, the context cap and the concurrency cap; those three numbers decide whether the box is enough.

Interview question

All weights set to 0: every neuron gets the same gradient, so the network never learns.

A whiteboard post asks whether you can initialise all weights to zero. Answer: no, the gradient is identical for every neuron, so weights must be random to break symmetry. A commenter adds that with no bias term everything stays exactly zero; another says use Glorot for sigmoid or tanh and He for ReLU.

the line to remember

Initialisation is about breaking symmetry and keeping signal variance stable through the layers. Zero does neither.

For your product

Nothing to decide here unless you train models. If you do, this is the first question a reviewer will ask about any training bug that ends with a network that outputs the same answer for everything.

Concept

3 agent patterns on one graphic: CodeAct, ReAct and agentic RAG. Two of the three descriptions oversell.

A graphic contrasts a single agent with a multi-agent system, then defines CodeAct (the agent acts by writing and running Python), ReAct (reasoning traces interleaved with tool actions, said to overcome hallucination and error propagation) and agentic RAG (agents orchestrating the retrieval pipeline).

the line to remember

ReAct is the loop, CodeAct is the action language, agentic RAG is the loop pointed at your documents. Pick the simplest one that fits the task.

For your product

Most business tasks do not need a multi-agent system. A single agent with a few well-described tools and a stop condition covers a lot; add agents only when the path cannot be known in advance.

Research

91.2 percent: how often innocent tool calls could be chained into a harmful action in Amazon's STAC study.

Two papers led by Amazon interns were accepted: one on relational priors in LLM multi-agent systems (AACL) and STAC, on how benign tools can form dangerous chains for LLM agents (EMNLP REALM workshop).

  • 91.2%mean final attack success (STAC)
  • 483generated attack chains

the line to remember

Agreement between agents is not accuracy, and per-tool safety is not chain safety. Evaluate the sequence, not the pieces.

For your product

If your agent can read, write and send, the danger is the combination. Approval gates on the final effect (money moved, mail sent, record deleted) matter more than filters on each tool.

Concept

32 parallel paths beat one wider block: ResNeXt's 'cardinality', doing the rounds again in 2026.

A walkthrough of 'Aggregated Residual Transformations for Deep Neural Networks' (Xie, Girshick, Dollár, Tu, He), the paper that introduced cardinality: repeating a block that aggregates a set of transformations with the same topology.

the line to remember

Split, transform, merge. The same idea now runs in mixture-of-experts language models: many small parallel paths instead of one big one.

For your product

A 2016 vision paper is still on the feed because the idea generalised. Grouped, parallel computation is why today's largest models can be cheap per token.

The daily dose, by day

Heard something about AI and not sure it is true?

Send it to us. We check it against the primary source and tell you what it means for your product.