Cesar Salem

It's inevitable.

[Cesar Salem @cesarsalem]

If you are working on a harness and it is making mistakes, don't go and fix the mistake that it makes, rather fix the harness so it can fix the mistakes.

[Cesar Salem @cesarsalem]

The progress we should be making MUST be always outcome/result-based

[Cesar Salem @cesarsalem]

LET'S GO!

[DeepSeek @deepseek_ai]

DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀

🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge.
🔹 On multimodal agent benchmarks, V4-Flash-Vision-Exp makes a major leap over V4-Flash, bringing multimodal agent performance close to Opus-4.8.

Try it with model='deepseek-v4-flash-vision-exp'. DeepSeek Harness 0.1.1 was released today with out-of-the-box support for the new model.

1/n

[Elon Musk @elonmusk]

One day, Optimus + Grok will provide incredible medical care to all the people of Earth

[Cesar Salem @cesarsalem]

I hope healthcare advancements like finding cures or developing drugs innovates LIKE AI model releases pace.

[Cesar Salem @cesarsalem]

Human taste is the hardest to automate.

[Cesar Salem @cesarsalem]

I love https://t.co/hn4w5DDkvi now :)

[Whop @whop]

Thrilled to partner with @spacexai to help people turn tokens into economic value.

Whop is now live as a native connector in @grok

[Cesar Salem @cesarsalem]

Compounding effect is getting real with Self-improving AI agents, solve or at least make an artificial human-like agentmemory + creating long-running loops that goes for days to weeks with cheap but powerful model to enhance the overall harness system you own probably the best move for now before applying them into any use-case you want.

Fail fast. More progress.

[Dario Amodei @DarioAmodei]

2/2 Second, on the messaging around AI.  I do not agree that my messaging has been disproportionately negative.  In fact it has been about equally balanced between risks and benefits: I’ve written one major essay about each, and even in interviews where I discuss the risks, I make sure to frequently mention the incredible benefits as well as proposing possible solutions to the risks (short clips from my interviews that end up on social media tend to be disproportionately negative, as that gets clicks).  In fact, I wrote Machines of Loving Grace because I didn’t feel the AI industry was painting an inspiring enough picture of how the technology could radically transform the world for the better.  The bulk of the essay is devoted to refuting skepticism of AI’s potential in health and biology, and showing why I think it will actually be possible to cure most human disease in ~5-10 years, as crazy as it may sound to ordinary people and frankly to biologists as well (I used to be one!).  And, if you read my most recent essay (Policy on the AI Exponential), I discuss concrete proposals for how to streamline the FDA process to make sure the deluge of AI-accelerated drugs isn’t slowed down by the regulatory process.  I feel the urgency here: I lost my father to Hepatitis C only a few years before the development of direct-acting antivirals (sofosbuvir), which cure 95% of patients and probably would have cured him.

I do agree that the public has a negative view of AI (and that this is a big problem), but I don’t think it is primarily caused by me or any other AI leader warning about AI’s risks.  I think it is fundamentally a crisis of trust.  I think that ordinary people don’t trust companies, governments, or the tech industry and always suspect that we are cooking up some new way to screw them over.  The causes of this go back decades and AI is just the latest iteration of it.  I don’t think that a glitzy marketing campaign with a positive spin (which some have advocated that Anthropic do) is the way to win back that trust — at this point, saying that AI will cure cancer is more a cliche than it is inspiring, and most people think it is deceptive.  The thing that will work is *actually curing cancer*.  I think by far the most accurate criticism of AI companies including Anthropic is that we haven’t yet delivered on our big promises to benefit the world.  That is totally on us, and I think it’s the criticism you should be making, instead of all this stuff about messaging and marketing.

We are however doing our best to fix this: Anthropic is ramping up its efforts very quickly in biology and medicine, and we hope to have incredible results in the coming years and some early glimmers in the coming months.  When we’ve actually accomplished something real, the whole world will hear about it, as loudly as possible, you have my word on that.  But until then I don’t want to make empty promises, and in the meantime I feel compelled to speak honestly about the very real risks of AI and how to address them.  Honesty is the right thing on the merits, and in terms of public credibility and trust it is no worse than, and may in fact be better than, an approach that ignores or distracts from risks which people instinctively understand are real.

[Cesar Salem @cesarsalem]

@hqmank @elonmusk Thanks for this, I thought Cursor Ultra will only last one-month!

[Shen Sean Chen @ShenSeanChen]

𝗜 𝗯𝗿𝗼𝗸𝗲 𝗱𝗼𝘄𝗻 𝘁𝗵𝗲 𝗲𝗻𝘁𝗶𝗿𝗲 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁 𝗺𝗲𝗺𝗼𝗿𝘆 𝗹𝗮𝘆𝗲𝗿: 𝗵𝗼𝘄 𝘁𝗼 𝘀𝘁𝗼𝗿𝗲 𝗶𝘁, 𝗳𝗶𝗻𝗱 𝗶𝘁, 𝗮𝗻𝗱 𝗸𝗲𝗲𝗽 𝗶𝘁 𝗰𝗹𝗲𝗮𝗻, 𝗶𝗻 𝗿𝗲𝗮𝗹 𝗰𝗼𝗱𝗲.

Free 30 min course on AI agent harness memories. Whiteboard first, then SQLite vs @mem0ai vs @zep_ai vs LangMem on the same facts.

🌟 Github Repo: https://t.co/dzS2f2Jdh1
💻 Join our Community: https://t.co/NMcEekLrmd

Most people bolt a vector database onto a chatbot and call it memory. The actual layer is four operations:
Store → Retrieve → Consolidate → Retire

@NousResearch Hermes and my own Waku Agent run zero embeddings. Plain text and one SQLite file. Open it, it's yours.

Retire is the one nobody implements. Deleting a fact and invalidating a fact are not the same system.

Full video in the first reply 👇
0:00 - Why every LLM call starts with amnesia
2:31 - Store memory as text, tables, or graphs
5:04 - Find it 4 ways: do nothing, keyword, RAG, Graph RAG
7:29 - Maintain it: add, delete, retire, attribute, reflect
9:47 - Build the whole thing in plain text and SQLite
13:13 - Add a vector store only when you actually need one
14:38 - Row memory vs graph memory vs temporal graph
21:03 - Run the code for all 5

You Can Build Anything. You Can Learn Anything. 💪

[Cesar Salem @cesarsalem]

@ShenSeanChen @mem0ai Now we're talking, here's how I will watch it later

[Cesar Salem @cesarsalem]

Tempting 😈

[Elon Musk @elonmusk]

I hope AI is nice to us

[Cesar Salem @cesarsalem]

Clearly a window for @MiniMax_AI to take its place.

[Wise @trikcode]

Chinese models were supposed to be cheaper

DeepSeek was supposed to kill OpenAI's pricing.

but starting tomorrow, DeepSeek V4 Flash output goes from $0.28 to $1.32 per million tokens at peak hours and then V4 Pro lands at $3.96.

I still can't believe this..

[Cesar Salem @cesarsalem]

Ouch.

[DeepSeek @deepseek_ai]

API pricing update 💰

With the V4 lineup release, we’re updating our API pricing and introducing peak and off-peak rates. Off-peak rates are 50% lower than peak, enabling more flexible workload scheduling. 📉

New pricing takes effect at 16:00 UTC, Aug 16, 2026 🕒

[DeepSeek @deepseek_ai]

🧩 DeepSeek Harness v0.1 is now available in Developer Preview!

🔹 We’re opening it up to developers building agent harnesses worldwide and open-sourcing the codebase in MIT license.
🔹 Powered by the Cordis meta-framework, DeepSeek Harness is an agent harness built around one core idea: Everything is a plugin. Models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and UI are ALL implemented as plugins, and can be mixed, matched, replaced, and extended.

Try it now!
https://t.co/2YWSvJHhKA

[Suraj Sharma @suraj_sharma14]

Advanced RAG techniques nobody's shipping.

Everyone demos naive RAG.
Almost nobody ships these.
This is the gap between a demo and a product.

1. Contextual Retrieval
Prepend each chunk with an LLM-generated context line before embedding.
Anthropic measured 35% fewer retrieval failures. 67% when combined with reranking.

2. Hybrid Search + RRF
Fuse BM25 (exact terms) with dense embeddings (semantics) via reciprocal rank fusion.
Catches queries that pure vector search misses.

3. Cross-Encoder Reranking
Retrieve 50 fast. Rerank down to 5 with a cross-encoder (BGE, Cohere).
Precision at top-5 jumps dramatically.

4. HyDE
Generate a hypothetical answer first. Embed that not the query.
Fixes vocabulary mismatch between user questions and your docs.

5. Query Decomposition
Split complex questions into sub-queries. Retrieve in parallel. Merge results.
Required for multi-hop questions naive RAG fails on.

6. Small-to-Big (Parent Document)
Embed small chunks for precision. Return the parent chunk for context.
Best of both: tight matching, full context.

7. Late Chunking
Run the whole document through a long-context encoder first. Then pool per chunk.
Every embedding carries global document context.

8. ColBERT / Late Interaction
Token-level matching instead of single-vector similarity.
Wins on technical, domain-specific queries.

9. Semantic Chunking
Split by meaning (embedding breakpoints), not fixed token counts.
Stops chunks from cutting ideas in half.

10. Contextual Compression
Extract only the relevant sentences from retrieved chunks before prompting.
Less noise. Fewer tokens. Higher accuracy.

11. Corrective RAG (CRAG)
Grade retrieved docs. If confidence is low, rewrite the query or hit web search.
Your system self-heals instead of hallucinating.

12. GraphRAG
Extract entities and relations. Build community summaries.
Answers "whole corpus" questions vector search cannot.

13. RAPTOR
Recursive clustering and summarization into a tree. Retrieve at multiple levels.
Multi-hop reasoning across the corpus.

14. Lost-in-the-Middle Ordering
Place strongest docs at the start and end of context. Attention is U-shaped.
Free accuracy gain from ordering alone.

15. Retrieval Evals
Track hit rate, MRR, NDCG on a golden dataset. Block deploys on regression.
You cannot improve retrieval you do not measure.

THE IMPLEMENTATION ORDER

Week 1: Semantic chunking + hybrid search

Week 2: Reranking + parent document retrieval

Week 3: Query decomposition + HyDE for hard queries

Week 4: Corrective loop + evals

Do not ship all 15 at once.
Stack them one at a time. Measure each.

Naive RAG is a demo. Advanced RAG is a product.

One retrieval pipeline plus public benchmarks plus documented tradeoffs
equals more credibility than 10 chatbot wrappers.

Most people ship the tutorial version. Builders ship the production version.

Bookmark & Repost!

[Cesar Salem @cesarsalem]

Thanks @XHandles!

[Cesar Salem @cesarsalem]

Is @ArtificialAnlys okay right now?

[Cesar Salem @cesarsalem]

Drowned to AI innovations now

[Artificial Analysis @ArtificialAnlys]

SpaceXAI's Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, joining the frontier in line with GPT-5.6 Sol, with standout agentic performance at lower cost

Grok 4.6 gains 5 points over Grok 4.5 on the Intelligence Index just over one month after its release, or +23 points compared to Grok 4.3. This brings SpaceXAI back to the intelligence frontier alongside OpenAI, behind only Anthropic.

Key takeaways:
➤ Grok 4.6 joins the frontier of the Artificial Analysis Intelligence Index: It scores 61, in line with GPT-5.6 Sol (max), behind Claude Opus 5 (max, 63) and Claude Fable 5 (max with fallback, 62), and just ahead of Kimi K3

➤ Strong agentic performance: Grok 4.6 achieves a GDPval-AA v2 Elo of 1753, behind only Claude Opus 5 and with overlapping confidence intervals with Claude Fable 5 and Qwen3.8 Max. It scores 50.7% on 𝜏³-Banking, among the top two scores alongside Qwen3.8 Max (51.3%), and 88.4% on Terminal-Bench v2.1, in line with the leading models

➤ Frontier-level intelligence at lower cost: Headline pricing is unchanged from Grok 4.5 at $2/$6 per 1M input/output tokens, 60%+ below Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30). It cost $0.84 per task, the same as Kimi K3 with slightly higher intelligence, placing it on the Intelligence vs. Cost per Task Pareto frontier

➤ Grok 4.6 sits at Fable 5-tier on AA-Briefcase, our private benchmark of long-horizon agentic knowledge work tasks, with an Elo of 1577 - behind the Claude Opus 5 family. It is notably turn-efficient, completing tasks in ~53 turns and ~0.5B input tokens on average vs. ~103 turns and ~2.0B input tokens for Claude Opus 5 (max)

Other model details:
➤ Context window of 500k tokens (unchanged from Grok 4.5)

➤ Pricing of $2/$6 per 1M tokens of input/output; cache hits discounted to $0.5 per 1M tokens, an increase over Grok 4.5’s $0.3 per 1M tokens for cache hits

Congratulations to @SpaceXAI and @elonmusk on the release!