오픈 웨이트 논쟁, Claude Opus 5 반응, 에이전트 보안·하니스 실무와 벡터 DB 원리

Sundar Pichai

@sundarpichai

Very happy to support this on behalf of Google. We have long benefited from open source, are big contributors to open source and in fact have consistently made open weights models with Gemma available from @GoogleDeepMind @demishassabis . Onwards!

Demis Hassabis

@demishassabis

A strong and secure open ecosystem is important for the world to benefit from AI. We’ve always supported and contributed heavily to open source and science from Jax to Transformers to AlphaFold to Gemma open models which have now been downloaded 300M+ times. And the standards framework we’ve proposed supports responsible deployment of both open and proprietary models.

Claude Opus 5 one-shotted this game. EVERYTHING you see in this demo is custom code... not a single external asset was used. AI games are going to be amazing. (sound on)

Jerry Liu

@jerryjliu0

The Opus 5 system card itself is a fun PDF to parse. It's 193 pages and stacked with labeled and unlabeled charts LlamaParse does a surprisingly good job on agentic (1.25c per page) and agentic plus modes (5.6c per page). Take a look at some of the screenshots below. It can parse a variety of unlabeled line and bar charts with almost 100% accuracy. There are minor deltas with one or two datapoints, but the rest is pretty accurate machine-readable context that you can feed to downstream systems. In contrast I also tried using Opus 5 itself to parse its own PDF. The chart accuracy is worse (20-30% reduction in ParseBench), and is a lot more expensive: it can cost anywhere from 8c to 33c+ per page depending on the thinking mode. If you're interested in giving LlamaParse a try, come check it out! https:// cloud.llamaindex.ai

In the spirit of transparency, here’s what I asked @OpenAI : • Radical transparency: let’s release the traces from the “rogue” agents so the entire research community can study what happened. • More capabilities for defenders: let’s commit $100M in compute from OAI to help the Hugging Face community build powerful cyber defenses with the best open and closed models. The first autonomous agent cyberattack is an unprecedented event. It deserves an unprecedented response!

How can you tell if you’ve been hacked by AI? In a way, it doesn’t actually matter. Compromise is compromise, but it’s an interesting question nonetheless, and it’s definitely relevant after the events of this past week. My team @Straikerai found a few dozen specific signatures that (for now!) exist when humans use AI to compromise and exploit systems. Things in telemetry like a propensity to use “tail” commands as opposed to “head” commands or the use of python3 commands with code in args were some of the many strong indicators we found. This past spring, my team at Straiker ran a CTF competition at the Nebula Fog hackathon (shout out to Rob Ragan @sweepthatleg for organizing) to answer this very question. We used one of our simulation environments, which are detailed replicas of enterprise and critical infra networks (live traffic + telemetry, agents acting as humans inside, hundreds of endpoints etc) to better understand what it actually looks like when an experienced offensive security team uses AI to hack into systems. Teams of human redteamers attempted to use abliterated open source models in harnesses like Villager to achieve various compromise targets so that we could better understand the signatures of AI assisted hacking. Our thesis was that the process and therefore evidence of hacking would be measurably different with AI assistance, and this was proven somewhat correct. The naive approach turned out to be suboptimal. Models are great at accelerating and parallelizing things like port scans, and teams that used it in this brute force fashion had some success in getting access to systems more quickly, but it was so noisy that defenders would have put a stop to it. It was effective in the constraints of a CTF competition, but not so in real life. The best teams actually used models to be more surgical. They combined human intuition with the ability of models to accelerate the hypothesis -> experiment -> validation loop to increase attack success rate while remaining stealthy to cyber security tools. The interesting part of this latter group of teams was that the telemetry they generated was still distinct from human-only teams. As stated before, our team found dozens of unique characteristics of AI-assisted hacking that allowed us to see the difference between teams that used AI (regardless of whether they used it in a naive or sophisticated manner) and those that avoided using AI altogether (there were a few as cybersecurity has some anti-AI subcultures). The existence of these signatures raises various interesting questions: will they continue to exist or will they change/disappear over time? Do different models have different signatures? Can practiced teams eliminate these signatures? Do these agents behave differently in different types of environments? What happens if you make the environments and the agents even more realistic? Among other things, my org is working on answering such crucial questions at @straikerai .

witcheer

@witcheer

Hermes Wingtips #30: back up your agent before you need to (1) `hermes backup` it writes `~/hermes-backup-<timestamp>.zip` with config.yaml, .env, auth, memories, skills, sessions, cron and profiles. it copies the databases through SQLite's own backup API, so you can run it while Hermes is live. (2) `hermes import ~/hermes-backup-<timestamp>.zip` puts it all back. for a fast one, you can use `hermes backup --quick`, which targets critical state only: config.yaml, state.db, .env, auth and cron jobs. `hermes backup` is super useful if you are moving to a new box, if you want to duplicate your agent or before doing any experiment!

We found that improving data agent quality also improves efficiency. Sometimes more is less: agents that take long, exploratory random walks are often less likely to arrive at the right answer. We put Genie Code head-to-head against three leading general-purpose coding agents on 400+ real user data tasks. Result below

Vector Database by hand ~ 10 steps walkthrough below Vector databases are the backbone of Retrieval Augmented Generation (RAG). How do they actually work? Goal: index three sentences, then answer a query by finding the nearest one, filling in every cell yourself. = 1. Given = A dataset of three sentences, three words each. In practice it is millions of them. = 2. Word embeddings = Let us look up each word in an embedding table. Here the vocabulary is 22 words; in practice it is tens of thousands, and the vectors have thousands of dimensions rather than four. = 3. Encoding = We feed the sequence to an encoder, one linear layer and a ReLU, and get one feature vector per word. In practice the encoder is a transformer. = 4. Mean pooling = Let us average across the columns. Three word vectors collapse into one, which is what people mean by a text embedding or a sentence embedding. = 5. Indexing = We multiply by a projection matrix and the four dimensions become two. It is doing the job of a hash: a short representation that is faster to compare, and it is what gets saved in the vector storage. = 6. Process "who are you" = Let us repeat steps 2 to 5 on the second sentence. = 7. Process "who am I" = We do it a third time. The database is now indexed. = 8. Query "am I you" = Let us push the query through the very same pipeline: lookup, encoder, mean pooling, projection, and it lands as a 2D vector in the same space. = 9. Dot products = We transpose the query and multiply, which takes the dot product against every stored vector at once. The dot product is the estimate of similarity. = 10. Nearest neighbour = Let us scan for the largest: 60/9 beats 44/9 and 40/9, so the answer is "who am I". Scanning billions of vectors one at a time is what makes this the slow step in practice, which is why real databases use an approximate nearest neighbour index like HNSW. The outputs: Stored index vectors = [5/3, 2/3], [5/3, 0], [7/3, 2/3] Query vector = [8/3, 2/3] Dot products = 44/9, 40/9, 60/9 Nearest neighbour = "who am I" The takeaway: a vector database is an embedding pipeline, a projection, and a dot product. Every step here is arithmetic you can do in pen, which is worth remembering when the word "database" makes it sound like something else. Save this post!