Daily Top Five

Today’s Top Five: Nov 8, 2026

By Willa Ravenwire · · 6 min

WIRE

Why it matters. Anthropic explains how agents should store and consolidate memory, plus IREN's Microsoft delay, Baidu's open ERNIE base, Perplexity and Meta.

The weekend was quiet on fresh announcements, so today’s list catches up on late week stories we had not covered yet, all checked against the companies’ own pages. Agents are the common thread: how they should remember, how they get sold to busy people, and how they start to talk back with a face. Underneath it all sits the bill for the computers, and one data center builder just showed how easily that schedule slips.

1. Anthropic Lays Out How Agents Should Remember

What happened. Anthropic’s engineering team published a guide to agent memory on Nov 6 that splits it into four layers: raw session transcripts, notes the agent writes, larger structures such as wikis, and a small block of context loaded into every session. It argues agents should also consolidate offline, a step it calls dreaming, where the agent rereads its logs to fix bad notes and reorganize what it knows. In one test, memory lifted Opus 5.5 on a database task from 0.655 to 0.838, and a wiki plus search answered 61% of questions on a small document set against 53% for search alone.

The hype check. The authors say several results come from single runs or tiny task sets, including a 25 question corpus and individual Pokemon playthroughs. The gains for older models were much smaller, a wiki barely helped when plain text search was enough, and a wrong note in the always loaded layer can spread errors into every session. Running the dreaming step also costs compute, and the post does not say how much.

Verdict: Promising but unverified. It is a useful map for anyone building agents, but the numbers are the vendor’s own early tests, not an independent benchmark.

2. IREN Pushes Back Its AI Cloud Target After a Customer Design Change

What happened. Data center operator IREN reported quarterly results on Nov 5 showing revenue of $228 million, up from $137 million, with $200 million of it from AI cloud work, and a net loss of $263 million. It moved its goal of more than $4 billion in annualized revenue from the end of 2026 to January, and possibly the end of February, blaming a design change the customer asked for plus supplier timing. The company also expects about $15 billion of further spending this fiscal year, with $9 billion covered and $6 billion still to be raised.

The hype check. The release names Microsoft as the customer for its Horizon 1 site, but it does not spell out which project the design change hit or what it costs IREN. The $20.3 billion backlog sounds solid, yet $6 billion of this year’s spending is not lined up, and the release itself says raising it depends on market conditions.

Verdict: News. Even contracted AI capacity is running late, and the funding gap is now on the record.

3. Baidu Releases Open Weights for a 280 Billion Parameter ERNIE Base

What happened. Baidu posted ERNIE 5.5 Flash Base on Hugging Face, a text only mixture of experts model with 280 billion total parameters and about 14 billion active for each token. Baidu says it trained the model from scratch on 33.5 trillion tokens, with a 128,000 token training context that stretches to one million tokens for fine tuning and inference. Its own comparison table shows it ahead of several other open base models on tests such as MMLU and HumanEval+.

The hype check. This is a raw pretrained checkpoint with no chat or reasoning training, so it is a starting point for builders, not a ready assistant. The benchmarks come from Baidu’s own evaluation pipeline, rivals still lead on several tests in that same table, and it ships under a custom Baidu license rather than a standard open source one, so read the terms before you build on it.

Verdict: Promising but unverified. Big open weights are real and downloadable today, but the scores need outside testing.

4. Perplexity Teammates Are Agents That Keep Working After You Leave

What happened. Perplexity launched Teammates, persistent agents that take ownership of ongoing jobs such as tracking metrics or sorting bug reports, each with its own running thread. They can plan, call in more agents, use a browser and desktop, and check back when they need a decision. Teammates are on the web now for Max subscribers, with phone apps listed as coming later.

The hype check. Perplexity’s page offers no data on how reliably these agents finish long tasks. Usage is free only during a promotion of unstated length, after which it draws on the plan’s monthly credits, and the page does not say how fast heavy work burns through them.

Verdict: Needs follow-up. Always on agents are where the industry is heading, but cost and reliability are still open questions.

5. Meta Gives Muse a Voice and an Animated Face

What happened. Meta began rolling out real time voice chat and expressive avatars in its Muse app on Nov 5, starting in the US on mobile, web and Mac. Users can choose from hundreds of voices, cut in while Muse is talking, and watch an avatar react as they speak. Meta says voice is coming to its AI Glasses and Muse Charm, with no date given.

The hype check. The announcement has no independent testing or latency figures, and it says nothing about pricing or other countries. Meta first showed the avatar technology at its Connect event in September, so this is a rollout of a known demo rather than a new capability.

Verdict: News. A talking, animated assistant is now a mainstream consumer feature, and the rest of the field will be measured against it.

Also on our radar

  • A robotics paper called VFlash decodes boxes and motion paths as whole shapes instead of one coordinate at a time, and reports planning several times faster on edge hardware.
  • Tencent’s WeChat team described WeLM-Image, one model for image generation and editing whose fast version makes a 1024 pixel image in about five seconds on a single GPU.
  • Researchers in China built SLA, a robot control model trained with 3D perception features, and report a real world success rate 24 points above the policy it started from.
  • A paper on training agents with sibling harnesses sends failed tasks to slightly different setups, then trains the model on the rescued attempts so it keeps those skills.

What to watch

See whether outside developers reproduce Anthropic’s memory gains in their own agents, and whether Baidu ships the chat tuned ERNIE 5.5 Flash with clearer license terms. On the money side, watch whether IREN closes its $6 billion funding gap and hits its new January target. For Perplexity and Meta, the test is real use: how much Teammates cost once the promotion ends, and how quickly Muse voice reaches glasses and countries beyond the US.

Share on X

Sources

  1. Anthropic: The Components of Agent Memory
  2. IREN via SEC EDGAR: IREN Reports Q1 FY27 Results (Form 8-K, Exhibit 99.1)
  3. Baidu on Hugging Face: ERNIE 5.5 Flash Base model card
  4. Perplexity: Teammates
  5. Meta: Give Your Muse a Voice
  6. Hugging Face Papers: VFlash, Fast, Accurate Visual Planning via Flash Primitive Decoding
  7. Hugging Face Papers: WeLM-Image, A Unified Image Generation Model towards Visual Intelligence
  8. Hugging Face Papers: SLA, Scaling Spatial-Language-Action Model via Continual Pre-training
  9. Hugging Face Papers: Harness-S, Learning from Sibling Harnesses for Generalizable Agent Training

By Willa Ravenwire · The AI Bandwagon

Next from the wireToday’s Top Five

More in Daily Top Five

  1. Daily Top Five · 6 min

    Today’s Top Five: Nov 22, 2026

    The Lean and Mathlib communities say Alpöge's machine checked proof of the Riemann hypothesis is correct, plus a canary study of leaky inference providers.

    Nov 22, 2026

  2. Daily Top Five · 6 min

    Today’s Top Five: Nov 21, 2026

    Anthropic closed its IPO with about $93.5 billion raised, as two rival Riemann hypothesis proofs and a new open model from Thinking Machines landed.

    Nov 21, 2026

  3. Daily Top Five · 6 min

    Today’s Top Five: Nov 20, 2026

    Anthropic opens on Nasdaq at $111.50, Claude takes on black hole physics, and Gemini learns to drive your Mac.

    Nov 20, 2026