AI Roundup August 2024: Grok-2 Ships FLUX, Anthropic Adds Prompt Caching, SB 1047 Clears the Legislature
TL;DR
August was an open-weights month. Black Forest Labs walked out of stealth and dropped FLUX.1, the first image model in a while that genuinely rivals the closed players, then immediately powered xAI's Grok-2 image generation. Anthropic shipped prompt caching and quietly handed builders a way to cut API bills by up to 90 percent. And California's SB 1047 passed the legislature, sending the most serious frontier-AI safety bill in the US straight to Governor Newsom's desk.
Black Forest Labs Walks Out of Stealth With FLUX.1, and It Actually Competes
On August 1, a new German lab called Black Forest Labs came out of stealth with FLUX.1, a family of text-to-image models that, for once, lives up to the launch hype. The team is largely ex-Stability AI: the people who built Stable Diffusion in the first place. They raised a 31 million dollar seed led by Andreessen Horowitz, which is small money in 2024 terms and a reminder that you do not need a nine-figure war chest to ship a frontier-grade image model.
The release came in three flavors: FLUX.1 [pro] (closed, API-only), FLUX.1 [dev] (open weights, non-commercial, guidance-distilled), and FLUX.1 [schnell] (Apache 2.0, fast, genuinely free to use). For the local-AI crowd this is the part that matters. Schnell runs on consumer hardware, the license is permissive, and the prompt adherence and hand/text rendering are a clear step up from the SDXL era. ComfyUI support landed almost instantly.
If you run a homelab image stack, FLUX.1 [schnell] is the new default to beat. Open weights, Apache 2.0, and quality that does not make you apologize for not paying for Midjourney.
xAI Ships Grok-2 and Grok-2 mini, With FLUX Doing the Image Generation
On August 14, xAI announced Grok-2 and Grok-2 mini in beta for X Premium and Premium Plus subscribers. xAI claimed the model was competitive with the top tier of the moment on reasoning, math, and graduate-level science benchmarks, putting it in the conversation with GPT-4 Turbo, Claude 3.5 Sonnet, and Llama 3.1 405B. Benchmarks are benchmarks, so take the leaderboard talk with the usual salt, but a credible third frontier-ish model arriving this fast is the real signal.
The headline drama was the image generation. Grok-2 wired in FLUX.1 from Black Forest Labs, and unlike most competitors it shipped with almost no guardrails. Within days the timeline filled with generated images of public figures and copyrighted characters doing things their owners did not authorize. It was a deliberate bet on minimal moderation, and it made for a sharp contrast with Google's still-paused Gemini people-generation. Love it or hate it, it put the open-versus-locked-down image debate front and center.
Anthropic Ships Prompt Caching, and Your API Bill Gets a Lot Smaller
On August 14, Anthropic launched prompt caching in beta on the Claude API for Claude 3.5 Sonnet and Claude 3 Haiku. The pitch is simple and the impact is real: cache a long, stable prefix (a big system prompt, a document, a codebase, a few-shot example set) once, then reuse it across calls. Anthropic quoted cost reductions of up to 90 percent on cached input and latency improvements of up to roughly 2x for long prompts.
This is one of those features that does not trend on social but quietly changes what you can afford to build. If you are stuffing the same 50k-token context into every request (RAG over a fixed corpus, an agent with a fat tool spec, a coding assistant that re-reads the same files), caching turns a budget-killer into something you can actually run in production.
Why builders should care
- Long system prompts get cheap: the fixed part of your prompt stops costing full price on every call.
- Agents and tool-heavy apps win most: static tool definitions and instructions are exactly the kind of repeated prefix caching is built for.
- It is a design constraint, not just a knob: structure your prompts with the stable stuff up front and the variable stuff last to get the most out of the cache.
California's SB 1047 Clears the Legislature and Heads to Newsom
The frontier-AI safety fight came to a head in Sacramento. SB 1047, the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act, was amended on August 15 (partly in response to feedback, including from Anthropic, which backed the revised version while flagging concerns). Then it moved fast: the Assembly passed it 48 to 16 on August 28, the Senate concurred on August 29, and the bill landed on Governor Newsom's desk.
The bill targets only the largest models, those trained above a compute and cost threshold, and would require safety testing, a shutdown capability, and accountability for catastrophic harms. OpenAI, Meta, Google, and a long list of industry groups lobbied hard against it, arguing it would chill open-source and hand an edge to less-regulated rivals. For anyone shipping open weights, the open-source carve-outs and the compute threshold were the clauses to watch. Newsom had until the end of September to sign or veto, and the whole industry was watching.
Key Takeaways
- Open weights had a real moment: FLUX.1 proved a small, focused team can ship a frontier-grade image model and hand the good parts to the community under a permissive license.
- The frontier is getting crowded: Grok-2 arriving this quickly means there are now several credible model families to choose from, which is good news for pricing and bad news for anyone betting on a single moat.
- Prompt caching is a cost lever, not a gimmick: if you run long fixed contexts, restructuring prompts around the cache can change your unit economics overnight.
- Minimal moderation is now a product strategy: Grok-2's near-guardrail-free image gen forced everyone to take a position on the open-versus-locked-down tradeoff.
- AI regulation got concrete: SB 1047 reaching Newsom's desk turned the abstract safety debate into a real bill with real compute thresholds that builders had to read carefully.