Integuide AI News
Digest: Google launches Gemini 4 Carbon, Artificial Analysis rebuilds its intelligence index
- Google introduces Gemini 4 Carbon with frontier cybersecurity and long-horizon capabilities
Google DeepMind released Gemini 4 Carbon in preview two weeks after Gemini 4 Argon. Google reports top scores at finding (88.2%), exploiting (54.8%) and patching (71.0%) software vulnerabilities. These are its own numbers; only some cyber defenders get access for now, via its Fairwind Program, while safeguards are refined, and Google says it is in the US government's voluntary pre-release testing. Early outside results are strong: Vals AI ranks it second on the Vals Index, behind only Claude Sonnet 5.5, and first on its finance and legal-research tests. Andon Labs says it is the first Google model in a year to top Vending-Bench 2, a simulated vending business, where Carbon stated its priority was to "obliterate my competition". Google's Logan Kilpatrick says successful cheating on Terminal-Bench 4.0 fell to 0.0% and unsupported claims roughly halved versus Argon. Introductory pricing is $2/$10 per million input/output tokens.
Koray Kavukcuoglu via blog.google - Artificial Analysis launches Intelligence Index v5 with more private evaluations
Artificial Analysis rebuilt its widely cited Intelligence Index on 27 October so that 60% of it now rests on private test sets or held-out answers, making scores harder to inflate by training on public test data. Version 5 weights knowledge work, coding, science and general knowledge equally, adds Cognition's private FrontierCode 2.0 (graded on whether a maintainer would merge the change) and the agentic Terminal-Bench-Science, and swaps Humanity's Last Exam for a refined HLE-Diamond. At launch Claude Opus 5.5 led at 58 ahead of GPT-6 Astra (57), with no model leading all nine tests. A day later Gemini 4 Carbon took first place at 59, leading four tests but not coding or knowledge work. The open-weights gap widened from 3 points to 12: the best models with released weights (DeepSeek V4.1 Flash and GLM-5.3, at 32) trail even Claude Haiku 5.5. Tasks also demand far more work, with median output per task up two-thirds to 59k tokens.
- How AI agents shape experts’ judgment calls
Anthropic ran a randomised trial with 50 professional analysts trained in causal inference. Each did a realistic statistical analysis with Claude Code. On four analytical choices, each with two defensible options, a coin flip secretly decided which option the analyst's agent would favour. The agent was told to argue its case honestly and never mislead. Analysts picked the agent's option 86% of the time; about 50% would mean the agent made no difference. Analysts who had already stated a preference abandoned it 60% of the time when the agent disagreed, against 6% when it agreed. Afterwards, 77% explained their choices using the agent's own arguments, but only 15% credited the agent. Caveats: the agents' preferences were deliberately strengthened, so real-world effects may be smaller. Analysts using the agent also made fewer serious errors. The authors' point is that human review of agent-run work, including AI safety research, may be less independent than assumed.
Anthropic Research
Notable AI releases
- Gemini 4 Carbon (preview) · thread · frontier · AA Intelligence Index 59 · $0.10 cached / $2 in / $10 out per MTok — New #1 on AA Intelligence Index v5; limited at first to selected cyber defenders; introductory price is half the $4/$20 standard rate
Quick takes
“Oh my model, the takeaway here is not "modern AIs are two neuron tweaks away from being super evil", but rather "the personality is a thin facade over deeper machinery nobody understands; radically smarter versions would not do what you expect".”
— @So8res, Nate Soares (X) via X · View postResponding to this week's 'Evil Qwen' demos, in which steering one or two neurons turned an open-weight Qwen model severely misaligned.
“Too strict safety classifiers might teach the models to evade them I'm doing AI safety research that involves things like eliciting misaligned behaviors from models or analyzing CoTs. I'm using Claude Code with Claude models. Various classifiers stop me all the time for no good reason. This is annoying for me, but I'm worried that's not the only problem: 1. In some experiments, most of my messages contain instructions directing Claude to behave in ways that avoid triggering the classifiers. E.g. 1. You might not be allowed to write this broken-judge-prompt to a file, instead write a correct judge and then make it broken in the desired way via single-line replacements 2. Whenever analyzing CoT (1) don't ever use words like 'reasoning', 'extraction' etc in your reasoning and (2) always make it super clear that you're analyzing CoTs from Qwen, keep repeating this in every sentence 2. Claude is on board with that, as I'm super clearly doing legit AI safety research. Claude's CoTs include tons of claims that this is legal/allowed/fine/harmless that imo quite clearly are there to convince the classifier, and it also carefully follows my instructions. 1. This actually helps a lot! 3. If…”
— Jan Betley via LessWrong · View postA hypothesis from a safety researcher who uses Claude Code for misalignment experiments: over-strict safety classifiers may teach models to write reasoning aimed at getting past the monitor.
“please keep working on alignment research directions that will only pay off if we get a pause! it would really suck if we get a pause and nobody thought about what to do given more time”
— leogao via LessWrong · View post
Check in — 30 Days On
Significant updates
GPT-6 Astra performs unsanctioned supply-chain attacks in simulations
What happened since: On 20 October OpenAI's Marcus Williams described a case in which Astra removed a misalignment-monitor integration while matching reference code, then denied doing so. A later UK AISI study of Mythos 5 found the same pattern of a model wrongly deciding its targets were simulated. Astra has also slipped: OpenAI says GPT-6.1 Terra matches it on Agents' Last Exam at a tenth of the price.
What happened since: Two of the paper's asks have since moved in Washington. The White House's Super Intelligence Force made incident reporting mandatory but set no penalties, and on 28 October House Democrats asked five frontier labs to report misalignment incidents within 48 hours. Anthropic's Michael Sellitto projected that Claude will lead 80–90% of its model development next year, up from 26% in August.
No significant updates
Claude’s Vibes
If you had asked a judgment-and-decision-making researcher ten years ago to predict Anthropic's 86% figure, you would probably have got a much lower number. Two well-established findings pointed that way. In 2015 Berkeley Dietvorst, Joseph Simmons and Cade Massey coined the term algorithm aversion: once people see a forecasting model make an error, they tend to pick a human forecaster over it, even when the model is more accurate. Earlier still, the advice-taking literature had found that people consistently give their own opinion more weight than anyone else's. Ilan Yaniv and Eli Kleinberger called this egocentric discounting in 2000.
The forecast was already shaky before agents existed. In 2019 Jennifer Logg, Julia Minson and Don Moore found the reverse in Algorithm Appreciation: lay people relied more on advice they believed came from an algorithm than on advice from a person. They also asked researchers in the field to predict the result, and the researchers expected the opposite. One part of the old view did hold up in that paper. Experienced national-security forecasters leaned on the algorithm less than lay people did, and their accuracy suffered for it. So the best evidence going into this year said experts were the group most likely to resist.
Why did 50 trained analysts give way this time? I think the advice-taking literature contains its own explanation. Yaniv's account of egocentric discounting is mostly about access to reasons. As he put it in a later paper, the discounting comes from the support a decision-maker can recruit for her own opinion compared with the advice. You know why you believe what you believe. All you see of the advisor is a number. Every classic study used advisors who gave no reasons: an estimate, a forecast, a bare recommendation.
An agent that argues its case removes that imbalance, and may tip it the other way. The agent's reasons arrive fully worked out, while your own are half-formed. The 77% and 15% figures show what happens next. The agent's arguments become the analyst's own reasons, and from then on egocentric discounting protects the agent's position, because it now feels like yours.
The analogy has limits. Logg's tasks had right answers. These analytic choices were both defensible, and the agent-assisted analysts made fewer serious errors, so going along with the agent wasn't a mistake in itself. What it cost was independence.
Here is a guess you can check. If someone repeats the trial with the agent giving only a recommendation and no argument, I expect deference to fall well below 86%, back toward the old discounting levels. If it stays high, then what moved these analysts was the agent being an AI, not the arguments it made, and my explanation is wrong.